Generative AI for Intelligent Transportation Systems: Road

Transportation Perspective
HUAN YAN, Tsinghua University, Beijing, China
YONG LI, Tsinghua University, Beijing, China

Intelligent transportation systems are vital for modern traffic management and optimization, greatly improv-
ing traffic efficiency and safety. With the rapid development of generative artificial intelligence (Generative
AI) technologies in areas like image generation and natural language processing, generative AI has also played
a crucial role in addressing key issues in intelligent transportation systems (ITS), such as data sparsity, dif-
ficulty in observing abnormal scenarios, and in modeling data uncertainty. In this review, we systematically
investigate the relevant literature on generative AI techniques in addressing key issues in different types of
tasks in ITS tailored specifically for road transportation. First, we introduce the principles of different gener-
ative AI techniques. Then, we classify tasks in ITS into four types: traffic perception, traffic prediction, traffic
simulation, and traffic decision-making. We systematically illustrate how generative AI techniques addresses
key issues in these four different types of tasks. Finally, we summarize the challenges faced in applying gen-
erative AI to ITS, and discuss future research directions based on different application scenarios.

CCS Concepts: • Applied computing → Transportation; • Computing methodologies → Machine

learning;

Additional Key Words and Phrases: Intelligent transportation system, generative AI, autonomous driving,
traffic flow

ACM Reference Format:

Huan Yan and Yong Li. 2025. Generative AI for Intelligent Transportation Systems: Road Transportation
Perspective. ACM Comput. Surv. 57, 12, Article 315 (July 2025), 45 pages. https://doi.org/10.1145/3719290

1 Introduction
The transportation system contains traffic infrastructure, equipment, and management strategies,
including traffic control and traffic planning, all aimed at managing traffic flow and meeting
travel demands. Its efficiency significantly affects societal and economic activities. Rapid urban-
ization has greatly increased the number of vehicles, leading to congestion and accidents. These
challenges inconvenience travelers and disrupt urban operations, emphasizing the need for better
traffic management. Recent advances in computing technology have enabled the development
of Intelligent Transportation Systems (ITS). Traditionally, ITS covers air, rail, road, and
marine transportation. In this article, we focus on urban road transportation and define ITS as an
integrated system connecting people, roads, and vehicles.

This work was supported in part by the National Natural Science Foundation of China under 62272260, U23B2030.
Authors’ Contact Information: Huan Yan, Tsinghua University, Beijing, China; e-mail: [email protected]; Yong Li
(Corresponding author), Tsinghua University, Beijing, China; e-mail: [email protected].

This work is licensed under a Creative Commons Attribution 4.0 International License.
© 2025 Copyright held by the owner/author(s).
ACM 0360-0300/2025/07-ART315
https://doi.org/10.1145/3719290

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:2 H. Yan and Y. Li

The foundational framework of ITS consists of four main components, including traffic
perception, prediction, simulation, and decision-making. Traffic perception collects and analyzes
large-scale data through sensors, cameras [277], and other monitoring devices to understand
traffic dynamics, which in turn supports prediction models for forecasting future demands and
conditions. Traffic simulation models traffic patterns under different scenarios, concentrating on
realistic traffic state distributions to validate decision-making feasibility [228, 327]. The decision-
making component leverages insights from perception, prediction, and simulation to implement
effective traffic control measures and optimize transportation operations, while also generating
new traffic data to evaluate their impact. Thus, these components form a cohesive closed-loop
system that addresses the majority of ITS applications, enabling effective management, control,
and optimization of transportation networks [146, 313].
Traditional deep learning methods have been widely adopted in ITS [152, 214, 217, 265, 275].
These methods can process a large number of traffic data, automatically recognize patterns,
predict trends, and optimize decisions, thereby improving traffic efficiency and safety. For
example, convolutional neural networks (CNNs) are used to analyze traffic camera images,
identifying vehicles, pedestrians, traffic signs, and road conditions to monitor traffic flow and
situations. Recurrent neural networks (RNNs) are effective for processing sequential data,
such as time series data from traffic sensors, enabling predictions of traffic flow, congestion,
and accident risk. Deep reinforcement learning (DRL) is applied to the field of autonomous
driving, where models are trained to make effective driving decisions by simulating the behavior
of intelligent agents, such as autonomous vehicles, in traffic environments. By utilizing these
deep learning techniques, ITS can more effectively analyze road conditions, predict congestion,
and provide adaptive decision support to optimize traffic efficiency.
However, with the increasing complexity of transportation networks, the integration of multiple
transportation modes, and the growing demand for travel have made the traffic environment more
complex. This complexity poses numerous challenges for traditional deep learning methods in
addressing traffic-related issues.
— Large amounts of sparse or low-quality data. Traffic data is often sparse and frequently suf-
fers from missing values or noise contamination. Traditional deep learning methods impose
high requirements on data quality, especially supervised learning approaches that typically
require a substantial amount of labeled input and output data for supervised training. The
scarcity or absence of traffic data can significantly impact model performance.
— Rare abnormal scenarios. Traffic environments exhibit high levels of dynamism, with occa-
sional occurrences of anomalies such as traffic accidents that pose significant threats to traf-
fic safety. These anomalies are challenging to capture as they are often rare events, resulting
in a lack of relevant samples in training data. Traditional deep learning methods are more
suitable for processing tasks under normal circumstances, and struggle to effectively address
rare or previously unseen abnormal situations.
— Unexplored modeling of uncertainty. In the transportation field, there are many factors that
may cause uncertainty. For example, adverse weather conditions such as heavy rain, snow, or
smog, as well as the uncertainty of various traffic participants’ behaviors such as changing
lanes and emergency braking, make traffic flow prediction more challenging. Traditional
deep learning methods usually adopt deterministic approaches, relying on pre-processed
features, thus making it difficult to effectively capture such uncertainty.
Generative Artificial Intelligence (Generative AI) technology has recently developed
rapidly in recent years, enabling the automatic generation of content, including text, images,
videos, and audio, based on user input or instructions. Several established techniques play key

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:3

roles in generative AI. For example, Variational Auto-Encoder (VAE) [114] is a generative
adversarial network (GAN) that models and generates complex data. GAN [79] improves the
quality of generated content through the adversarial process of generators and discriminators.
Other techniques include normalizing flow [194] for generating complex data distributions,
energy-based models [122] for enhancing generation quality, and generative models from
physical processes [158] for generating content with physical constraints. Additionally, diffusion
probabilistic models [210] generate data by adding noise to data gradually to reduce its quality.
Generative Pre-trained Transformer (GPT) [74] use pre-trained Transformer architectures
[224] to generate text.
These generative AI techniques offer significant advantages in data generation, reconstruction,
and data uncertainty modeling, which hold the potential for addressing challenges in ITS [161].
For example, in autonomous driving scenarios, generative AI can generate high-fidelity driving
scene images and videos, which are crucial for training and testing autonomous systems [117, 120].
By using AI-generated simulation scenarios, autonomous vehicles can perform extensive driving
simulations in virtual environments, improving their ability to make accurate real-world decisions.
In traffic flow prediction tasks, generative AI technology can learn and model the distribution of
traffic data, thus generating future traffic flow data [245, 284, 291]. This capability helps traffic
authorities improve traffic planning and optimization.
Given the growing potential of generative AI in ITS, it is crucial to conduct an extensive review
of past research that focuses on generative AI techniques applied to ITS. As shown in Table 1,
numerous surveys have been conducted on ITS or generative AI. However, most of these studies
do not delve deeply into the practical applications of generative AI in ITS. For example, while one
study focuses on GANs [146] and another on diffusion models [184] in ITS, neither comprehen-
sively addresses other generative AI techniques. Moreover, these surveys do not systematically ex-
amine how generative AI can overcome challenges faced by traditional deep learning approaches
in ITS. To bridge this gap, this article explores the applications of generative AI techniques in traf-
fic perception, prediction, simulation, and decision-making within ITS, as shown in Figure 1. Our
goal is to provide a comprehensive analysis of their role in promoting intelligent transportation,
and offer a constructive insight to help readers gain a deeper understanding of the challenges and
opportunities of generative AI in ITS. We expect this survey to provide useful insights for further
research and advancement of ITS.
Our survey focused on high-quality literature from reputable conferences, books, and journals.
We referred to established archives like arXiv and Google Scholar, and major databases such as
IEEE Xplore, ACM Digital Library, Springer, and ScienceDirect. Keywords like “Intelligent Trans-
portation System” combined with generative AI terms (e.g., “variational autoencoder,” “generative
adversarial network,” “GPT,” “LLM”) were used to identify relevant references. Additional filters
included terms like “traffic signal,” “traffic prediction,” “traffic simulation,” and “data imputation.”
We screened literature by examining titles, eliminating low-quality papers based on citation count,
author affiliation, and publication authority, followed by an abstract review to assess contributions.
This process ensured a foundation of credible, high-quality research.
In conclusion, the primary contributions of this article are illustrated as follows:
— To the best of our knowledge, we are the first to offer an in-depth literature review of gen-
erative AI for ITS.
— We provide a systematic introduction to mainstream generative AI techniques, conduct in-
depth method comparisons from both horizontal and vertical perspectives, and provide a
systematical analysis of how generative AI technology can effectively address key issues in
ITS from the aspects of traffic perception, traffic prediction, traffic simulation, and traffic
decision-making.
ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:4 H. Yan and Y. Li

Table 1. Related Reviews of ITS and Generative AI

Category Reference Description
Comprehensive [6, 64, 68, 72, 77, 90, 166, 189, 192, 195, 206, 209, 221, These works give a
254, 287] comprehensive review of ITS,
covering key technologies,
infrastructure, and applications.
However, few generative AI
techniques are systematically
introduced.
Technologies DRL [92], GNN [190], NLP [186], SVM [88], DL These works conduct surveys
[83, 225], Positioning [59, 98], GAN [146], Diffusion on specific techniques within
models [184], Blockchain [50, 218], Edge the field of in ITS.
computing [78, 322], Dynamic pricing [200], Fuzzy
logic [218], Swarm intelligence [169], V2X
downloading [237], Data mining [7, 60, 325],
Communication [20, 52, 112, 164, 168, 173, 205]
Infrastructure IoT [15, 23, 182], VANET [13, 222], Bluetooth [70], These works conduct surveys
ITS Radar-on-chip [202], Hardware devices [49], on specific components in ITS.
Distributed architecture [176], External
infrastructure [44]
Applications Security management [2, 3, 5, 14, 89, 121], Traffic These surveys focus on
scheduling [175], Traffic detection [150], Driver specified applications in ITS.
behavior detection [38], Vehicle classification
[75, 272], Metaverse [178], Urban mobility [167],
Urban monitoring [145], Traffic sign detection
[61, 160], Traffic prediction [213, 279], Traffic
simulation [25], Autonomous driving [4, 47], Trust
management [163], Spectrum regulation [41]
Comprehensive [21, 69, 247, 289] These works conduct a
comprehensive review of
Generative generative AI techniques.
AI Applications 3D [126], Biology [296], ChatGPT [240, 288], These surveys concentrate on
Metaverse [187], Edge cloud computing [241], the specific applications of
Mobile network [57, 260], Security & Privacy generative AI techniques in a
[24, 238], Vehicle network [297] particular field.
DRL: Deep reinforcement learning, GNN: Graph neural network, NLP: Natural language processing, V2X: Vehicle to
everything. DL: Deep learning. IoT: Internet of things. VANET: Vehicle ad-hoc network.

— We discuss the open challenges encountered in applying generative AI technology in ITS,

and explore potential directions for future research.
The structure of this article is as follows. Section 2 introduces mainstream generative AI tech-
niques and gives a detailed comparative analysis. The following four sections respectively discuss
the applications of generative AI technology in traffic perception, traffic prediction, traffic sim-
ulation, and traffic decision-making. Finally, in Section 7, we introduce the challenges faced by
generative AI in the application of ITS, and look forward to future research directions.

2 Fundamental Generative AI Technologies

As described in previous works [21, 69, 240, 247], generative AI refers to a subset of AI algorithms
that create personalized and high-quality content such as text, images, videos, and 3D assets based
on user input or specific requirements. This section provides an overview of the fundamental
technologies behind generative AI.

2.1 Key Technologies

With the rapid advancement of computer technology and computing power, generative AI has
made significant progress over the past decade. In 2013, Kingma and Welling introduced VAE

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:5

Fig. 1. An overview of generative AI in intelligent transportation systems.

Fig. 2. The history of generative AI techniques.

[114], a probabilistic graphical model designed for data generation. In 2014, Ian Goodfellow
et al. designed GANs [80], which had a profound impact on the field of generative AI. In
2018, OpenAI’s GPT made significant contributions to the field. Beyond these, technologies
like Normalizing Flow, Energy-Based Models (EBMs), Diffusion Probabilistic Models
(DPMs) and Generative Models from Physical Process (GenPhys) have shown substantial
potential for text and image generation. Figure 2 depicts the evolution timeline of generative AI
techniques, including VAE, GANs, Normalizing Flow, EBMs, DPMs, GenPhys, and GPT. In this
subsection, we will provide an initial introduction to these techniques, followed by a comparative
analysis.
2.1.1 VAE. Variational Autoencoders (VAEs) [114] aim at learning latent representations
of data and generate new samples by mapping data to a latent space and using a decoder to
reconstruct it.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:6 H. Yan and Y. Li

Given observed data x from an unknown distribution pθ (x), VAEs∫ introduce latent variables z
and model pθ (x|z). The data distribution is represented as: pθ (x) = z p(z)pθ (x|z) dz. Since direct
computation is intractable, VAEs use an encoder qϕ (z|x) to approximate the posterior pθ (z|x),
minimizing the Kullback–Leibler (KL) divergence between these distributions.
In ITS applications, this approach allows capturing latent patterns in traffic data, generating
synthetic traffic data that can be used for simulation, traffic prediction, and traffic decision-making.
2.1.2 GAN. GANs [80] consist of a generator and a discriminator, trained together in an
adversarial manner to create realistic data samples. The generator creates fake data from random
noise, while the discriminator tries to distinguish it from real data The goal is to reach a point
where the generator produces data that the discriminator cannot reliably differentiate from real
data, achieving a Nash equilibrium. This adversarial process enables ITS researchers to generate
high-quality synthetic traffic scenarios that can be used for various applications, e.g., vehicle
control. Generative Adversarial Imitation Learning (GAIL) extends GANs to learn policies
by imitating expert behaviors without requiring explicit reward signals [93].
2.1.3 Normalizing Flows. Normalizing flows [194] transforms a simple probability distribution
into a complex one by applying a series of invertible transformations. These transformations are
reversible, allowing for efficient calculation of probability densities.
Let f be an invertible function that maps a random variable z, sampled from a simple distribution
∂f −1
p(z), to a new variable y = f (z). The probability density of y is given by: p(y) = p(z)| det ∂y | =
∂f
p(z)| det ∂z | −1 . By applying K transformations to an initial variable z0 ∼ p0 (z0 ), we obtain a final
K ∂f
variable zK and its log probability density: ln pK (zK ) = ln p0 (z0 ) − k=1 ln | det ∂zkk−1 |.
This allows starting from a simple distribution, like a Gaussian, and creating more complex,
multi-modal distributions. In the context of ITS, this allows modeling complex traffic data dis-
tributions while maintaining tractable probability densities. By applying normalizing flows to a
simple distribution, we can accurately estimate the likelihood of specific traffic conditions, which
is crucial for tasks like anomaly detection and probabilistic traffic forecasting.
2.1.4 Energy-Based Models. EBMs [58, 122] assign energy values to data points, with lower
energy indicating a better fit. The goal is to assign lower energy to realistic points and higher
energy to unrealistic ones, approximating the likelihood of observed data.
−E (x)
The probability density for data x is defined using a Boltzmann distribution: pθ (x) = ∫ e e θ−Eθ (x̃) ,
x̃∈X
where Eθ (x) is an energy function parameterized by θ . To optimize EBMs, contrastive divergence
compares positive samples (from real data) with negative samples (generated by the model), updat-
ing parameters to reduce energy for positives and increase energy for negatives. This is repeated
iteratively to improve the model. Due to their flexibility and robustness, EBMs are used to model
complex traffic patterns and perform multivariate time series forecasting in ITS applications [266].
2.1.5 Diffusion Probabilistic Model. DPMs [210] generate data by reversing a noisy process to
recover the original data step by step.
In the forward process, a Markov chain successively adds Gaussian noise to the data x0 , resulting

in an approximate posterior: q(x1:T | x0 ) = Tt=1 q(xt | xt −1 ).
The reverse process starts from a noise distribution p(xT ) = N (0, I), and iteratively removes

noise using learnable Gaussian transitions: pθ (x0:T ) = p(xT ) t1=T pθ (xt −1 | xt ).
During training, the model predicts the noise added at each step, minimizing the difference
between the predicted and actual noise. In the ITS domain, DPMs are utilized to model and
generate complex traffic patterns by progressively denoising noisy traffic data. This enables the

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:7

creation of synthetic traffic scenarios, which are valuable for traffic simulation, prediction and
decision-making.
2.1.6 Generative Models from Physical Process. GenPhys [158] transforms physical partial
differential equations (PDEs) into generative models, generalizing frameworks like diffusion
models.
The goal is to generate new samples from observed data pdat a (x). A continuous physical process
∂p(x,t )
is described by: ddtx = v(x, t), which evolves the probability distribution p(x, t) as: ∂t + ∇ ·
[p(x, t)v(x, t)] − R(x, t) = 0, where R(x, t) represents particle birth or death. The goal is to design
p(x, t), v(x, t) and R(x, t) such that p(x, 0) = pdata (x) and pprior (x) ≡ p(x,T ) becomes independent
of pdata (x) as T → ∞. To generate samples, the reverse physical process is used to evolve x(T ),
drawn from p(x,T ), backward to produce x(0).
By incorporating GenPhys into ITS, we leverage the power of physics-informed generative mod-
eling to improve traffic simulations, leading to better decision-making and more efficient trans-
portation systems.
2.1.7 Generative Pre-Trained Transformer. GPT is a neural network designed for natural
language processing (NLP) and text generation tasks. GPT uses a transformer decoder archi-
tecture [224] with self-attention and feedforward layers to capture long-range dependencies in
text. It processes inputs with positional encodings and word embeddings, and generates text
in an autoregressive manner, token by token. Pre-trained on large text datasets, GPT can be
fine-tuned for specific tasks. These advanced capabilities can be leveraged to enhance various
aspects of ITS, particularly in autonomous driving. For example, by fine-tuning GPT models on
transportation-specific data, they can better understand complex driving scenarios, improving
the decision-making processes of autonomous vehicles.
2.1.8 Comparative Analysis of Generative AI Techniques. In previous subsections, we introduced
several major generative AI techniques, each with distinct advantages and limitations regarding
performance, stability, flexibility, and computational requirements. For example, VAEs effectively
learn latent representations suitable for data with inherent uncertainty, with moderate compu-
tational demands and stable training, but may struggle to generate high-quality images. GANs
generate high-quality, diverse images using an adversarial process but face challenges like unsta-
ble training, mode collapse, and high computational demands. Table 2 summarizes these aspects
for different generative AI techniques.

2.2 Important Generation Tasks

The advancement of generative technologies has provided impressive solutions for various gener-
ative tasks. In this section, we will explore the applications of generative technologies in ITS from
the perspectives of text, images, videos, and cross-modal data, as shown in Figure 3.
2.2.1 Text Generation. Text generation focuses on using AI techniques to autonomously gen-
erate reasonable and meaningful text. In ITS, text generation is employed for tasks such as gener-
ating real-time traffic reports and automated incident alerts [81]. This capability provides better
insight into traffic conditions and enables timely responses to dynamic traffic situations, thereby
improving overall traffic management and safety.
2.2.2 Image Generation. Image generation involves using AI algorithms to create new images
or visual content based on given information. In the context of ITS, image generation is utilized for
generating synthetic images for testing the driving algorithms in autonomous vehicles [140]. By
generating realistic images reflecting various traffic scenarios, weather conditions, and lighting

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:8 H. Yan and Y. Li

Table 2. Comparison of Generative AI Technologies Evaluated in Terms of Performance, Stability,

Flexibility, and Computational Requirements
Technology Performance Stability Flexibility Computational
Requirements
• Learns effective latent • Relatively stable • Capable of • Moderate computational
Variational representations training process probabilistic resources needed
Autoencoder [114] • Suitable for data with generation
inherent uncertainty
• May produce blurry images
• Generates high-quality and • Training can be • High flexibility in • Requires substantial
Generative diverse images unstable generation computational power
Adversarial • Prone to mode processes
Network [79] collapse
• Accurately models complex • Stable but • Supports both • High computational and
Normalizing probability distributions computationally generation and parameter requirements
Flow [194] intensive inference for high- dimensional
data
• Models complex distributions • Training can be • Supports both • High computational
Energy-Based effectively slow and requires generation and demands during training
Model [122] • Robust to outliers careful tuning inference

Diffusion • Robust to noise and • Stable training • Flexible noise • Longer generation time
Probabilistic uncertainty process schedule and complex training
Model [210] • Allows controlled generation
Generative Models • Generates data grounded in • Stability depends • Offers high • Requires prior
from Physical physical processes on the physical interpretability knowledge of physical
Process [158] model accuracy parameters
• Excels in natural language • Generally stable • Capable of • Requires substantial
Generative processing tasks once trained generating computing resources for
Pre-trained • Effective in modeling continuous and pre-training
Transformer [74] long-term dependencies coherent text

Partly referenced by the work [297].

Fig. 3. The relation between different generation tasks in ITS.

environments, generative methods help improve the robustness and performance of driving
behavior decision-making models used in ITS.

2.2.3 Video Generation. Video generation involves creating a series of frames that are semanti-
cally coherent to form a video. This is particularly important in ITS for simulating traffic scenarios,
training autonomous driving systems, and analyzing vehicle behaviors. The ability to generate
realistic traffic videos allows researchers and engineers to test and validate ITS algorithms under
a wide range of conditions without the need for extensive real-world data collection [311].

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:9

Fig. 4. The relation between different topics in generative AI for traffic perception research.

2.2.4 Cross-Modality Data Generation. Cross-modality data generation refers to generating

data in one modality based on features from a different modality. In ITS, this is crucial for tasks such
as converting textual descriptions of traffic situations into visual representations (text-to-image),
generating textual reports from traffic camera images (image-to-text), or creating simulations of
traffic scenarios based on textual inputs [199]. These techniques enable more comprehensive traf-
fic analysis and support the development of robust autonomous driving algorithms by providing
diverse and rich datasets.

3 Generative AI for Traffic Perception

Traffic perception refers to the ability of ITS to collect and understand sensory information from
the traffic environment. This information includes visual data from cameras, trajectory data from
GPS-based devices, motion data from accelerometers, and environmental data such as weather
and road conditions, as well as other traffic-related information like the presence of other vehi-
cles, pedestrians, and road signs. Traffic perception is essential for making effective decisions and
ensuring safety. In autonomous vehicles, traffic perception typically relies on cameras, LiDAR,
millimeter-wave radar, and other technologies to collect and process data from the vehicle’s sur-
roundings, enabling autonomous driving and navigation.
However, accurate perception and comprehension of the complex dynamics within traffic envi-
ronments present several significant challenges. First, sensors may fail to provide data in certain
situations, such as when cameras are obstructed. These data gaps can disrupt the continuous mon-
itoring of traffic conditions, potentially leading to incomplete information. Second, sensor data
collected in traffic environments is subject to various sources of noise, including sensor inaccu-
racies, weather conditions, and lighting variations. This noise introduces uncertainty, making it
harder to derive precise information. Third, traffic environments are inherently complex and dy-
namic, involving numerous interacting factors such as vehicles, pedestrians, traffic signals, and
road conditions, which accurately modeling and understanding traffic behavior challenging.
Recent advancements in generative AI technology offer promising solutions to these challenges.
We carried an in-depth literature review focusing on the application of generative AI to address
challenges in traffic perception. The reviewed literature covers various aspects of traffic percep-
tion, including data imputation, traffic estimation, data analysis, and anomaly detection. Their
relationships are illustrated in Figure 4.

3.1 Traffic Data Imputation

Traffic data imputation involves learning a function, f (·), that takes corrupted or missing data
X as input to reconstruct the complete data X̂ . This can be expressed as X̂ = f (X , Ψ), where Ψ
represents additional factors like the road network graph.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:10 H. Yan and Y. Li

Low-dimensional traffic data. Low-dimensional traffic data, such as speed, flow, and travel
time, is crucial for various applications like route planning and traffic prediction. However, missing
data is common due to sensor failures or transmission issues. To address this, researchers have
employed advanced generative AI techniques to impute these missing data.
VAE. Boquet et al. [18] developed an online unsupervised VAE-based method for imputing miss-
ing traffic data, enhancing forecasting accuracy while enabling data augmentation, compression,
and classification. Building on this, Chen et al. [28] introduced the STVAE model, which incorpo-
rates 3D gated convolutions, a multi-attention network, and Sylvester normalizing flows to better
capture complex spatio-temporal dependencies.
GAN. Recently, more researchers have adopted GAN-based methods to address this task
[33, 86, 110]. For example, Chen et al. [33] proposed using real or corrupted data as latent codes
and introducing representation loss within a GAN model. IGANI [110] introduced a novel iterative
GAN imputation architecture, which iterated over imputed data while preserving the reversibil-
ity of the generative imputer. In particular, numerous works proposed to incorporate the spatio-
temporal modeling in GAN-based architectures [94, 95, 129, 131, 185, 207, 267, 268, 300, 300, 301].
For example, the authors in [142] introduced a 3D convolutional GAN that excels in capturing
intricate traffic dynamics through fractionally strided and standard 3D CNNs for the generator
and discriminator, respectively. STGAN [281] enhanced GANs with generation and center losses,
along with skip connections and dilated convolutions, achieving superior imputation while main-
taining data consistency. GAE-GAN-LSTM [259] integrated a graph autoencoder with GANs and
LSTM, offering robust imputation across various loss ratios by effectively modeling complex traf-
fic patterns. These innovations highlight the superior capability of advanced GAN-based methods
in capturing and reconstructing the complex spatio-temporal relationships in traffic data, offering
more reliable and versatile solutions.
DPM. Recently, diffusion models have attracted more researchers’ attentions in traffic data im-
putation [157, 282, 299]. For instance, Liu et al. [157] introduced PriSTI, an enhanced conditional
diffusion framework that handles diverse missing patterns and high missing rates by integrat-
ing conditional feature extraction and advanced noise estimation. This method mitigates error
accumulation seen in autoregressive models and incorporates geographic relationships, offering
a more context-aware imputation. Similarly, SaSDim [299] enhanced diffusion models with a self-
adaptive noise scaling mechanism and a spatial-temporal global convolution module. Their novel
loss function normalizes noise intensity, reducing the negative impact of noise and capturing dy-
namic dependencies more effectively. These methods illustrate the potential of diffusion models to
address spatio-temporal imputation tasks by leveraging their inherent flexibility in handling noise
and capturing complex data dependencies.
Finally, to provide a clear illustration of the spatio-temporal modeling within the generative AI
framework for this task, we summarize the related works in Table 3.
High-dimensional traffic data. High-dimensional traffic data, such as traffic videos and
3D LiDAR vision plays a crucial role in enhancing situational awareness and decision-making,
particularly in autonomous driving. However, the loss of high-dimensional data remains a
common challenge in real-world applications. To address this, Wu et al. [248, 249] utilized
feature pyramid networks within GAN frameworks to extract multi-scale semantic features
from traffic videos. This multi-scale approach enhanced the integration of spatial and temporal
information, ensuring continuity and accuracy in the imputed frames. The incorporation of
local and dual-branch discriminators further guaranteed spatio-temporal consistency, which
was crucial for applications requiring precise motion and context understanding, such as traffic
monitoring and autonomous navigation. Tu et al. [223] extended GAN applications to 3D LiDAR
point clouds with their Point Fractal Network (PF-Net). By repairing incomplete point clouds,

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:11

Table 3. Spatio-Temporal Modeling in Generative AI Methods for Low-Dimensional Traffic Data

Imputation

Method Spatio-Temporal Modeling Papers

MLP [18]
VAE
3D gated CNN, multi-attention mechanism [28]
3D CNN [142]
Dilated CNN [281]
CNN [268, 293]
CNN, self-attention mechanism [301]
Temporal: Gramian Angular Summation Field [94]
Temporal: discrete Wavelet transform [95]
GAN Temporal: multi-view temporal factorizations [131]
Spatial: learnable bidirectional attention map; Temporal: multi-channel matrix [267]
Spatial: GCN; Temporal: LSTM [207, 259]
Spatial: multi-layer FCN; Temporal: LSTM [129]
Spatial: dynamic GCN; Temporal: multi-head self-attention network [185]
Spatial: spatial attention GCN; Temporal: self-attention GRU [300]
Spatial: GNN [282]
DPM Spatial: GCN; Temporal: Transformer [157]
Spatial: dynamic GCN; Temporal: global temporal convolution [299]

PF-Net restores vital geometric and semantic information, thereby improving object detection
and collision avoidance in autonomous vehicles. This GAN-based repair not only enhances data
integrity but also supports the real-time requirements of autonomous driving systems.

3.2 Traffic Estimation

Traffic estimation aims to learn an estimator to estimate the current traffic condition of all roads
in the transportation network. This task differs from traffic data imputation in that it addresses
the challenge of incomplete data collection due to constraints such as cost and time limitations.
Several works designed GAN-based approaches to estimate traffic condition [143, 172, 220, 262,
278, 308, 309]. GAA [143] merged traffic-flow theory with GANs using LSTM networks to effec-
tively capture spatio-temporal correlations. This integration not only improved estimation accu-
racy compared with Bayesian methods but also demonstrated the potential of combining theo-
retical models with deep learning. GCGA [278] combined GCN with GANs, capturing the spatial
dependencies and leading to more accurate traffic condition estimations. Innovations like Curb-
GAN [306] and GE-GAN [257] further refined these models by incorporating dynamic convolu-
tional layers and graph embeddings, enabling more precise and context-aware traffic estimations.
PA-GAN [220] integrated Bayesian inference to adaptively adjust parameters based on contextual
features, effectively handling multi-modal traffic patterns and reducing estimation biases. Addi-
tionally, Zhang et al. [309] models the complex correlations between road traffic and urban condi-
tions in their proposed conditional GAN-based model, further improving estimation accuracy.
Unlike them, the authors in Ref. [280] leveraged denoising DPM by parameterizing noise fac-
tors, transforming traffic matrix estimation into a gradient-descent optimization task. This method
demonstrated superior performance in both traffic matrix synthesis and estimation. Lei et al. [124]
further extended diffusion models with their Conditional Diffusion Framework with Spatio-
Temporal Estimator (CDSTE), which integrates probabilistic diffusion with spatio-temporal net-
works. CDSTE excels in inferring complete traffic states at sensor-free locations, providing reliable
probabilistic estimates that support robust traffic management and control decisions.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:12 H. Yan and Y. Li

Table 4. Related Works about Traffic Data Mining

Category Task Method Papers

Transportation mode identification VAE [302]
Trajectory
Decision-making behavior analysis GAN [304]
Image transformation GAN [140]
Image denoise GAN [316]
Image resolution enhancement GAN [16, 97]
Vehicle classification GAN [239]
Vehicle detection GAN [91]
Traffic image
License plate recognition GAN [85, 123]
Traffic density identification GAN [73]
Behavior detection GAN [246]
Road annotation GPT [106]
Flow inference Normalizing Flow [318]
Vehicle-road user interaction detection VAE [36]
Object segmentation GAN [183]
Traffic video Activity classification GAN [118]
BEV perception DPM [328]
Accident video understanding DPM [62]

3.3 Traffic Data Mining

In this section, we introduce the related works that utilize generative AI techniques to analyze
traffic data from various sources, including trajectories, traffic images, and traffic videos, as sum-
marized in Table 4.
Trajectories. Trajectories are sequences of timestamped positions that describe the movement
of an object over time, providing valuable insights for analyzing various aspects such as urban
transportation modes and human decision-making preferences. To identify the transportation
modes, the authors in Ref. [302] introduced a semi-supervised Dirichlet VAE that integrates geo-
graphic data with motion features. This model effectively leveraged limited labeled GPS data and
accounts for geographic factors, achieving superior classification accuracy on real-world datasets.
Similarly, TrajGAIL [304] used Variable Length Markov Decision Processes and deep neural
networks to capture long-term dependencies in human decisions. This model outperformed
traditional Markov-based models, demonstrating enhanced accuracy in replicating complex
behaviors.
Traffic images. In autonomous driving scenarios, traffic images captured by cameras provide
essential information about the surrounding environment, allowing autonomous vehicles to
make effective decisions. To overcome the challenge of limited image diversity in real urban
environments, Li et al. [140] introduced a GAN-based framework that performs diverse urban
image style transfers, such as day-to-night transitions, while effectively preserving foreground
objects and structural details. This method outperformed baseline models by eliminating dis-
tortions and maintaining high-quality transformations, thereby increasing the variability and
reliability of training data for autonomous systems. Zhao et al. [316] addressed image degradation
from raindrops on in-vehicle cameras with a GAN-based removal network that integrates visual
attention and content perception. This approach not only enhances image quality but also
improves traffic object detection accuracy.
Unlike them, some researchers focus on extracting valuable information from traffic images
for tasks such as vehicle detection [239], license plate identification [16, 85, 97, 123], traffic

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:13

density recognition [73], driving behaviors [246] and map enrichment [318]. These tasks often
face challenges related to low resolution and inadequate lighting conditions. For example, the
authors in [91] enhanced vehicle detection in low-light conditions by integrating Pix2PixGAN and
CycleGAN with a YOLOv4-based algorithm. This method effectively overcomes the limitations
of traditional vision techniques in poor lighting, demonstrating substantial performance im-
provements that enhance autonomous vehicle reliability. In the domain of license plate detection,
Boby et al. [16] compared a GAN-based super-resolution model with diffusion-based iterative
refinement models. This comparison highlighted the importance of selecting the appropriate
super-resolution technique based on specific application requirements, with GANs offering
significant advantages in measurable performance metrics. In addition, the authors in Ref. [73]
presented a dual-discriminator conditional GAN for recognizing traffic density in both homoge-
neous and heterogeneous traffic scenarios. The dual-discriminator architecture, optimized using
Giza pyramids construction, enhanced the model’s ability to distinguish between varying traffic
densities with improved accuracy and reduced loss.
Traffic videos. Compared with traffic images, traffic videos provide a wealth of redundant traf-
fic information for autonomous vehicles, guaranteeing both safety and efficiency in navigation. A
pivotal stage in numerous intelligent video processing applications, such as vehicle behavior analy-
sis, is the segmentation of moving objects. RMS-GAN [183] is an end-to-end GAN using a recurrent
technique, which integrates foreground probability information using residual and weight-sharing
techniques to achieve precise segmentation. Krishna et al. [118] progressively combined genera-
tive and discriminative models to efficiently classify activities in traffic videos. Cheng et al. [36]
proposed a deep conditional generative model to detect interactions between vehicles and vul-
nerable road users. This model utilized a conditional VAE-based framework with Gaussian latent
variables to capture the behavior of road users and predict interactions in a probabilistic manner.
DiffBEV [328] addressed the issue of noisy Bird’s Eye View (BEV) representations in autonomous
driving by using a conditional diffusion model to improve BEV feature quality through progressive
refinement. The authors in Ref. [62] proposed an abductive framework AdVersa-SD for accident
video comprehension. AdVersa-SD employs an object-centric video diffusion approach to learn
cause-effect relationships of accidents, thereby enhancing safe driving perception.

3.4 Traffic Anomaly Detection

Traffic anomalies refer to unusual events, patterns, or behaviors within a transportation sys-
tem. These anomalies can disrupt the regular traffic flow or pose potential safety risks. Nu-
merous studies have explored the use of generative AI techniques for traffic anomaly detection
[119, 149, 188, 251]. Chen et al. [31] presented a multi-modal GAN model to detect and classify the
traffic events, leveraging data from multiple modalities to enhance detection accuracy. The authors
in Ref. [201] proposed a hybrid CNN-VAE architecture that employs color gradient representations
and semi-supervised annotation, achieving a 1–6% increase in classification accuracy and a 30–35%
boost in anomaly detection through effective feature extraction and semi-supervised labeling. The
work by Kang et al. [109] addressed congestion anomaly detection by integrating conditional nor-
malizing flows with kernel density estimators to effectively capture spatial-temporal dependencies
in multivariate time series data. Their framework not only demonstrated superior performance, but
also ensured robustness across diverse traffic scenarios. A-VAE [9] integrated 2D CNN and BiLSTM
layers with an attention mechanism for enhanced representation learning in traffic video anom-
aly detection. This approach demonstrated strong real-time processing capabilities, highlighting
its practical applicability for live traffic monitoring systems. Li et al. [125] presented DiffTAD, a
diffusion model-based framework for vehicle trajectory anomaly detection that addresses training
instability and mode collapse issues prevalent in GANs and VAEs. By modeling anomaly detection

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:14 H. Yan and Y. Li

as a noisy-to-normal process and incorporating Transformer-based encoders for temporal and

spatial dependencies, DiffTAD achieved superior performance and robustness in diverse traffic
environments. Lastly, Yan et al. [264] introduced a novel diffusion model–based method for video
anomaly detection that utilizes two denoising diffusion implicit modules to predict and refine the
motion and appearance features of video frames. The proposed method effectively learned the
distribution of normal samples, enabling accurate detection of anomalies.
Cyberattacks would pose significant safety risks in the context of automated driving. To
address it, researchers have proposed a variety of solutions leveraging generative models. Zhao
et al. [315] presented GVIDS, a GAN-based intrusion detection system for connected vehicles
that distinguishes legitimate CAN messages from malicious ones. Although GVIDS is suitable for
dynamic vehicular environments, its reliance on specific CAN message characteristics may limit
adaptability to new attack types. Ding et al. [55] exposed the vulnerability of deep generative
models in autonomous driving to Trojan attacks through data poisoning. Their work underscored
the critical risk of maliciously implanted behaviors that activate under specific triggers, highlight-
ing the need for robust data integrity measures. Li et al. [137] developed a GAN-based real-time
anomaly detection model capable of identifying subtle adversarial manipulations in vehicle and
sensor data, demonstrating its potential for dynamic environments. In addition, the authors
in Ref. [54] presented an unsupervised GAN-based framework optimized with LSTM networks
to analyze vehicle dynamics, effectively detecting eleven distinct attack types. By integrating
LSTM architectures, the framework effectively handled extensive vehicular data, demonstrating
robustness in maintaining the integrity of connected and autonomous vehicle networks.

3.5 Discussion
We have summarized the key insights and challenges of generative AI in traffic perception.
Generative AI techniques such as VAE, GAN, and DPM have shown advantages in handling
missing low- and high-dimensional traffic data, particularly in capturing complex spatiotemporal
dependencies, improving traffic prediction and management. However, their ability to handle
data noise and dynamic changes varies, requiring scenario-specific selection. Generative AI
has also enhanced traffic estimation by integrating traditional traffic theory and deep learning,
significantly improving accuracy. In traffic data mining, progress has been made in trajectory
analysis, image processing, and video analysis, but challenges related to data diversity and model
generalizability exist. In anomaly detection, generative AI effectively identifies both regular
and anomalous patterns, but real-time performance and adaptability need improvement. Future
research should focus on optimizing spatiotemporal modeling, improving data robustness, and
integrating integration with traditional traffic theory to optimize ITS.

4 Generative AI for Traffic Prediction

Traffic prediction refers to estimating or forecasting future traffic conditions, encompassing var-
ious aspects such as travel demand, travel time, traffic flow, and the movements and behaviors of
both vehicles and individuals within the transportation system. Traffic prediction is vital for urban
traffic planning and management. However, achieving accurate predictions presents significant
challenges. First, urban traffic systems exhibit complex spatial and temporal dynamics. Second,
traffic conditions are highly dynamic, susceptible to rapid changes due to factors like real-time in-
cidents, ongoing construction activities, and even human behavior. Third, obtaining high-quality
data from various sensors, cameras, and GPS devices is important for traffic prediction, but such
data is often limited and incomplete. Generative AI technology offers powerful solutions to over-
come these challenges. In our comprehensive literature review, we concentrate on the utilization
of generative AI in addressing traffic prediction challenges. We categorize these works into four

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:15

Fig. 5. The relation between different topics in generative AI for traffic prediction research.

main aspects: human-related, vehicle-related, road segment-related and region-related traffic

prediction, encompassing the key components of transportation systems. Their relation is given
in Figure 5.

4.1 Human-Related Traffic Prediction

Human-related traffic prediction involves forecasting individual behavior and human mobility.
Individual behavior. We discuss individual behavior prediction based on their roles within
transportation systems. For drivers, modeling and predicting driver behavior play a crucial
role in enhancing safety and preventing risky actions [12, 100]. For example, Bao et al. [12]
introduced a probabilistic sequence-to-sequence approach utilizing a conditional VAE to predict
various driving behaviors. For pedestrians, accurately forecasting human motion behavior is vital
for autonomous vehicles to plan preemptive actions and avoid collisions, ensuring pedestrian
safety [48, 216]. For instance, in [48], the authors proposed a temporal convolutional GAN for
high-fidelity future pose prediction, incorporating two discriminators: one for fidelity assessment
and another for consistency evaluation in long sequences. UTD-PTP [219] introduces a univer-
sal transformer diffusion modeling framework that combines transformer architectures with
diffusion-based uncertainty modeling to accurately predict pedestrian trajectories in complex
autonomous driving environments.
Human mobility. Human mobility refers to an individual’s movement from one location to
another [227, 229]. Feng et al. [67] presented a graph-based imitation learning framework that
adeptly models courier decision-making in route selection. By utilizing a multi-graph neural net-
work, ILRoute effectively captured the complex, multi-source features influencing route choices,
while incorporating mobility regularity and personalized constraints to navigate the expansive
decision-making space. This approach improved prediction accuracy across multiple real-world
datasets and online evaluations. Building on the necessity to understand and predict human
mobility patterns during unprecedented events like the pandemic, Bao et al. [11] introduced
a spatio-temporal conditional GAN Covid-GAN to estimate mobility changes under varying
public health conditions and policy interventions. Furthermore, the authors extended their initial
model to Covid-GAN+ [10] addressed spatial heterogeneity by using the local Moran statistic
and outliers by redesigning the training objective to learn the estimated mobility changes. This
method achieved a more accurate and robust estimation of human mobility.

4.2 Vehicle-Related Traffic Prediction

In the reviewed literature, vehicle-related traffic prediction tasks encompass vehicle trajectory
prediction and crash risk prediction.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:16 H. Yan and Y. Li

Vehicle trajectory. A vehicle trajectory refers to the path a vehicle takes from one location to
another over a specific time period. Precise vehicle trajectory prediction is vital for a variety of
traffic-related applications, including traffic management, navigation, and route planning.
Most problems of vehicle trajectory prediction are defined as using historical trajectories to
infer future trajectories. Formally, given a vehicle i with a historical trajectory xi = {r it } for t =
1, 2, . . . ,T , where r it represents the location at time t, the goal is to learn a mapping function д(·)
to predict the future trajectory yi = {r it } for t = T + 1, . . . ,T + n.
However, achieving precise trajectory predictions is a challenging task for several reasons. First,
vehicle movements are influenced by various factors, including driver behavior, traffic conditions,
road infrastructure, and unexpected events. Second, traffic conditions can change rapidly due to
incidents, congestion, or accidents. Third, the accuracy of trajectory prediction models heavily
relies on the quality and quantity of input data, such as GPS data. Noisy or incomplete data can
lead to inaccurate predictions. To address these challenges, more researchers proposed different
methods based on generative AI technology to predict vehicle trajectory.
VAE. Several researchers have designed prediction models based on VAE [37, 113, 179, 252].
DVAE [177] integrated expert knowledge into the decoder component of the VAE, which can
provide the similar prediction accuracy and enhance interpretability. The authors in [101, 102]
utilized conditional VAEs to encode past trajectories of vehicles and traffic scenes for vehicle path
prediction. By conditioning the network on these inputs, their approach effectively captured the
multimodal nature of vehicle movements in various traffic scenarios. Gui et al. [82] incorporated
additional dynamic information like yaw angle and velocity in their proposed conditional VAE
framework. This approach allows for a more comprehensive understanding of vehicle interactions
and movements, resulting in superior performance on both established and newly introduced sim-
ulated datasets.
GAN. GAN is also an alternative technique for trajectory prediction [139, 198, 321]. For exam-
ple, Kang et al. [107] applied social GAN, originally designed for pedestrian trajectory prediction,
for the prediction of vehicle trajectories. While this approach demonstrated better prediction ac-
curacy, its reliance on models initially tailored for pedestrian dynamics may limit its adaptability
to the distinct behaviors exhibited by vehicles. Differently, Zhou et al. [321] introduced the self-
attention social GAN to handle long trajectory sequences and complex vehicle interactions. The
self-attention mechanism helped capture correlations in trajectory data, reducing the loss of impor-
tant information over time. However, the model’s complexity may limit its suitability for real-time
applications requiring high computational efficiency. KI-GAN [243] was a knowledge-informed
GAN that integrated traffic signal data and multi-vehicle interactions through a specialized atten-
tion pooling mechanism. This model effectively incorporated orientation and proximity informa-
tion for accurate predictions but relied heavily on the quality of input data, which can vary across
urban environments.
GPT. Recently, GPT models have great potential in vehicle trajectory prediction [66, 212, 226].
Su et al. [212] proposed a crossmodal transformer based generative framework that integrates
multiple modalities and pedestrian attributes to predict pedestrian trajectories more accurately.
Similarly, Feng et al. [66] developed a pre-trained large traffic model using the Transformer ar-
chitecture. This model has the ability to capture the diverse trajectories within a population of
vehicles, leading to higher accuracy in vehicle trajectory prediction. However, the high computa-
tional demands of such large model may pose challenges for real-time deployment in autonomous
systems.
DPM. Diffusion models also attract more researchers’ attentions. For instance, EquiDiff [29] com-
bined the conditional diffusion model with an equivariant transformer, leveraging the geometric
properties of position coordinates and social interactions among vehicles for accurate prediction.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:17

Crash risk. Crash risk prediction is important in traffic safety management and precaution.
Due to imbalanced data between crash and non-crash cases, accurately predicting crash risk
is challenging. To deal with this, Cai et al. [19] proposed a deep convolutional GAN model to
generate synthetic crash data. This approach effectively preserves the heterogeneity of non-crash
scenarios, enabling the use of the entire non-crash dataset for training predictive models. Similarly,
in Ref. [165], Wasserstein GAN (WGAN) was developed to address the extreme class imbalance.
By generating realistic synthetic crash instances, WGAN achieved higher sensitivity in crash
detection while maintaining a low false alarm rate. Different from approaches using GAN, Zhang
et al. [290] proposed an enhanced accelerated testing method based on Normalizing Flows. This
method efficiently learned the distribution of rare crash events and preserves the correlations
between scenario variables, significantly reducing the number of required tests by over two
hundred times compared with traditional Monte Carlo simulations. However, the method’s
effectiveness is highly dependent on the quality of the learned distributions and may require
extensive computational resources for training the Normalizing Flows models.

4.3 Road Segment-Related Traffic Prediction

Road segment-related traffic prediction focuses on forecasting traffic conditions (e.g., traffic flow,
speed, travel time) for specific segments or sections of roads within a transportation network
[127, 128, 235]. Accurate prediction faces three main challenges. First, traffic conditions within
the same road segment can vary over time, but there is often similarity between adjacent time pe-
riods. Additionally, the traffic state of one road segment can affect neighboring segments, resulting
in complex spatial and temporal patterns in traffic conditions. Second, the limited number of road
sensors and the possibility of sensor malfunctions can result in missing or inaccurate data, pos-
ing challenges for ensuring data quality and completeness. Third, external factors like accidents
and weather further enhance the uncertainty of prediction. Generative AI technology has gained
significant attention from scholars due to its ability to learn data distributions and model uncer-
tainties. Numerous models based on generative AI technology are proposed for predicting traffic
flow and travel time.
Traffic flow prediction. The task of traffic flow prediction is to learn a non-linear function

h(·) from previous T time steps of traffic conditions to forecast traffic conditions for the next T
steps based on road network graph G. Many scholars employ different generative AI techniques
to tackle this task. We summarize the related works in Table 5.
VAE. Some researchers have proposed VAE-based models. The authors in Ref. [17] designed
a VAE-based prediction model with the ability to learn the data distribution, resulting in im-
proved forecasting performance and versatility across multiple real-world datasets. PFVAE [104]
presented a novel planar flow-based VAE for traffic flow prediction. By integrating LSTM as the
autoencoder and employing planar flows within the VAE structure, the model effectively captured
the intricate temporal dependencies and nonlinear patterns of time-series data. Empirical results
demonstrated its robustness and accuracy in handling noisy time-series data.
GAN. GAN-based models are commonly adopted in this task [53, 148, 203, 294, 310]. For
example, ForGAN [115] employed GAN to learn the underlying data distribution for accurate
probabilistic forecasts. This approach not only enhanced prediction accuracy but also provided
a robust framework for handling real-world fluctuations in time-series data. In particular, many
researchers designed several methods to model spatiotemporal correlations in GAN-based pre-
diction model [103, 111, 135, 256, 276, 285, 312, 317]. For example, PL-WGAN [103] utilized GCN,
RNN and attention mechanisms within a WGAN framework to model spatiotemporal correlations
in urban traffic networks, leading to highly accurate short-term traffic speed predictions. However,
the combination of multiple neural network components increased computational demands. Sim-

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:18 H. Yan and Y. Li

Table 5. Related Works about Road Segment-Related Traffic Prediction

Task Method Spatio-temporal modeling Papers

VAE Temporal: LSTM [104]
Temporal: LSTM [256]
Spatial: GCN [276, 317]
Spatial: GCN, Temporal: RNN [103]
GAN Spatial: GCN, Temporal: GRU and self-attention [111]
Traffic flow
prediction Spatial: GCN, Temporal: TCN [135]
Saptial: GraphSAGE; Temporal: LSTM [312]
Spatial: GCN; Temporal: attention network [53]
Normalizing flows Autoregressive network with convolution operations [284]
Spatial: GCN, Temporal: TCN [245]
DPM
Temporal: RNN [191]
Travel time GAN Spatial: GCN [211]
estimation DPM Masked Vision Transformer [151]

ilarly, ASTGAN [156] used the attention mechanism and a mask graph convolutional recurrent
network to characterize temporal and spatial dependencies, improving data completeness and
prediction accuracy. The GAN component ensured the reliability of forecasts, showing significant
improvements in real-world traffic prediction tasks.
Normalizing flows. Consider the capability of conditional normalizing flows in representing com-
plex distributions characterized by high dimensionality and strong interdimensional relationships.
MotionFlow [284] utilized conditional normalizing flows to model complex spatio-temporal data
distributions. This model effectively captured the intricate variability inherent in traffic patterns
through its autoregressive conditioning on spatio-temporal inputs.
DPM. Since DPM can effectively handle uncertainties within spatio-temporal graph neural
network (STGNN), DiffSTG [245] combined the spatiotemporal learning capability of STGNN
with the uncertainty measurement of DPM, resulting in higher prediction accuracy. The authors
in Ref. [191] developed an autoregressive denoising diffusion model for multivariate probabilistic
time series forecasting, utilizing gradient estimation and Langevin sampling to accurately sample
data distribution. Experimental results showed that it outperformed traditional and contemporary
forecasting models in both accuracy and robustness.
Travel time estimation. Travel time estimation task aims to estimate the duration it takes to
travel between two locations or along a specific route given the departure time. Several scholars
employ generative AI techniques to address this task [151, 211]. To be specific, GCGTTE [211]
integrated GCN and GAN to estimate travel time in a probabilistic form. This approach not only
enhances prediction accuracy but also provides valuable insights into traffic variability, which is
essential for proactive traffic management. Lin et al. [151] proposed a two-stage diffusion-based
origin-destination (OD) travel time estimation model. The model incorporated a conditioned
pixelated trajectory denoiser to capture correlations between OD pairs and their historical tra-
jectories, effectively managing variability and noise in trajectory data. It outperformed baseline
methods in terms of accuracy, scalability, and explainability.

4.4 Region-Specific Traffic Prediction

Region-specific traffic prediction focuses on forecasting traffic conditions such as traffic flow
within specific geographic regions in a transportation network. The main challenges in achiev-
ing accurate predictions include complex spatio-temporal dependencies, limited data, and noise.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:19

Recent works [65, 130, 174, 292, 305, 307] have aimed to employ generative AI techniques to tackle
these challenges. To be specific, Kang et al. [108] developed a GAN-based model that integrates
gated and dilated convolutions to effectively capture both local and distant traffic patterns within
urban regions. This dual convolution approach allows for more precise real-time traffic flow pre-
dictions compared with traditional grid-based models. Li et al. [134] proposed a seq2seq spatial-
temporal semantic GAN model for multi-step urban flow prediction. By treating successive urban
flow data as video frames, the model employed a spatial-temporal semantic encoder to simultane-
ously capture semantic factors and spatial-temporal dependencies. The integration of adversarial
loss with prediction error effectively mitigates the issue of blurry predictions, resulting in more
accurate and reliable forecasts. The authors in Ref. [174] incorporated recurrent and conventional
network models into the GAN framework to predict the taxi demand in specific areas. The model
demonstrated superior performance on real-world data, and provided valuable insights into the
dynamics of taxi supply and demand.

4.5 Discussion
Generative AI has made great progress in traffic prediction, addressing challenges in data quality,
behavioral modeling, and spatiotemporal dynamics. Despite their high accuracy, these models of-
ten lack interpretability, which is crucial for building trust in ITS. Approaches like DVAE [177],
which incorporate expert knowledge, improve interpretability without sacrificing performance.
This balance is crucial for gaining trust and facilitating the adoption of generative AI technology in
ITS. Real-time applications also present computational challenges. While GPT-based frameworks
[66, 212, 226] offer high prediction accuracy, they struggle with latency and resource demands,
highlighting the need for optimized models that balance performance and efficiency. In summary,
achieving a balance between accuracy, interpretability, and efficiency remains critical in traffic pre-
diction. Future research should focus on developing generative AI models that not only achieve
high accuracy but also maintain transparency and efficiency in real-time traffic environments.

5 Generative AI for Traffic Simulation

Traffic simulation involves the mathematical modeling of transportation systems, which can be
used to generate the movement and behavior of vehicles and pedestrians. This technology can
be applied in various fields, including urban planning, evaluation of policy changes, and testing
and validating autonomous vehicle algorithms. However, achieving a more realistic traffic simu-
lation poses several challenges. First, real-world traffic scene data is typically collected through
road cameras or sensors. However, obtaining such data is not only costly but also challenging to
capture in rare or hazardous situations. Second, traffic dynamics are inherently complex due to the
numerous variables involved, ranging from individual driver behaviors, decisions, and reactions
to the unpredictable nature of external factors such as weather or sudden events. Capturing the
intricacies of these dynamics is a considerable challenge. Additionally, the interconnected relation-
ships between vehicles, infrastructure, and pedestrians further complicate the situation. Existing
works using traditional deep learning methods [84, 250] often require extensive labeled data, and
may not generalize well to unseen or rare situations. In addition, these methods may struggle to
capture the vast array of dynamic interactions in traffic scenarios [132, 244].
Generative AI techniques offer potential solutions for traffic simulation challenges. They can
create realistic traffic scenarios, including rare or extreme situations that might be missing from
real-world data. Their adaptability allows them to adjust to ever-changing traffic dynamics, thereby
capturing intricate interactions more holistically. Furthermore, by learning from a small number of
real traffic data, they can refine their simulations continuously, ensuring they mirror the complex-
ities of real-world traffic more closely. Next, we will introduce the related works using generative

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:20 H. Yan and Y. Li

Fig. 6. The relation between different topics in generative AI for traffic simulation research.

AI techniques from three aspects: driver behavior simulation, traffic scenario generation and traffic
flow generation. We draw their relation in Figure 6.

5.1 Driver Behavior Simulation

In autonomous driving scenarios, the focus is on how drivers make decisions in response to traffic
conditions [26, 76, 208, 274, 283]. Yun et al. [283] utilized GAN to develop a data-driven, human-
like driver model capable of simulating intricate interactions in traffic scenarios. By leveraging
GANs, their model effectively captured the intricate driving skills exhibited by humans in diverse
and complex situations, outperforming traditional rule-based models. Chen et al. [26] introduced
a two-stage hybrid driver model that integrates data-driven neural networks with model-based
controllers. By employing reward-augmented GAIL, their model effectively simulated human-like
driving behaviors while ensuring safety and expressiveness in multi-agent highway scenarios. In a
novel application of LLM, Jin et al. [105] developed a generative driver agent simulation framework
that leverages human driver experience through self-reported driving thoughts. By integrating
LLMs with human-derived chain-of-thought data, their framework significantly reduces collision
rates and enhances the human-like qualities of simulated drivers. Addressing the computational
challenges associated with large-scale simulations, Ye et al. [274] proposed an efficient calibration
method for agent-based traffic simulations using VAEs. By compressing agent state vectors into
lower-dimensional latent spaces, their approach dramatically reduced the computational burden
of calculating state transfer probabilities.
Lane changing and car following are two common driving behaviors, and numerous studies have
explored methods for simulating these behaviors [56, 136, 147, 162]. Ma et al. [162] introduced a
physics-informed conditional GAN that integrates the advantages of both physics-based models
and deep learning techniques. This hybrid approach eliminated the need for explicit weighting
parameters, simplifying the model while effectively capturing multi-step car-following behaviors
in complex traffic environments. Lin et al. [147] introduced a constrained-GAIL framework that
employs reward augmentation to guide driving agents away from undesirable behaviors. By in-
corporating manually designed reward functions alongside those learned from data, their model
achieved higher prediction accuracy in vehicle speed and location while significantly reducing
collisions and erratic driving behaviors. Addressing the need for highly realistic simulation en-
vironments, the authors in Ref. [56] leveraged the transformer and diffusion models to generate
human-like lane-changing trajectories. Extensive experiments validated its ability to produce re-
alistic and controllable trajectories.
Unlike autonomous driving scenarios, in transportation services, a variety of factors such as
estimated travel demand tend to dominate drivers’ behaviors. For example, Zhang et al. [303]
developed a conditional GAIL model capable of learning the driver’s decision-making tendencies
and strategies through knowledge transfer across different drivers and locations. By integrating

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:21

Table 6. Related Works about Traffic Scenario Generation

Task Method Controllability Papers

VAE Unconditional [32, 295]
GAN Unconditional [42, 120, 144, 144, 180]
Vehicle trajectory DPM Unconditional [215, 255, 261, 326]
generation VAE, GAN Unconditional [116, 117]
GAN Conditional [8]
DPM Conditional [319, 320]
Human trajectory GAN Unconditional [253]
generation DPM Conditional [193]
VAE Unconditional [19, 99]
Anomaly data
GAN Unconditional [30, 35, 96]
generation
GAN Conditional [286]

collective inverse RL, cGAIL enhances the generalization of driver behaviors, thereby improving
traffic management and service quality.

5.2 Traffic Scenario Generation

Traffic scenario generation refers to produce specific traffic situations or anomalous events for the
purpose of testing and validating autonomous driving systems. As autonomous vehicles need to
operate safely in a vast array of driving conditions, it is crucial for developers to test these systems
against a wide range of potential real-world scenarios, especially those that are rare or potentially
hazardous. In this literature review, we will focus on three key areas: generating vehicle trajecto-
ries, generating human trajectories, and generating anomaly data, as summarized in Table 6.
Vehicle trajectory generation. Many researchers use generative AI technology to generate
more vehicle trajectory data [32, 42, 116, 117, 120, 144, 180, 215, 255, 295]. In Ref. [117], the au-
thors introduced trajectory GAN and trajectory VAE to generate realistic synthetic lane-change
maneuver trajectories without expert knowledge. This approach not only facilitated the creation
of diverse training datasets for autonomous driving systems but also reduced the dependency on
extensive expert annotations. PS-TrajGAIL [232] modeled urban vehicle trajectory generation as a
partially observable Markov decision process within a GAIL-based structure. This approach lever-
aged IL to capture the complex decision-making processes of human drivers, thereby generating
synthetic trajectories that exhibited high fidelity to real-world behaviors. Experimental evalua-
tions on both synthetic and real-world datasets demonstrated that PS-TrajGAIL significantly out-
performs existing baselines in terms of trajectory realism and predictive accuracy. Diff-Traj [326]
was designed for generating GPS trajectories by learning spatio-temporal features from histori-
cal trajectories. By employing DPMs, this model captured the stochastic and dynamic nature of
human driving behaviors, producing high-quality synthetic trajectories that retain the statistical
properties of real-world data. WcDT [270] combined DPMs with transformers to generate diverse
and realistic autonomous driving trajectories. By encoding historical data and environmental fea-
tures through transformer-based encoders, WcDT effectively captured multifaceted interactions
and temporal dependencies present in driving scenarios. The integration of diffusion models en-
hanced the framework’s ability to produce high-fidelity trajectories that reflect the variability and
unpredictability of human driving behavior.
Differently, some searchers focused on the controlled vehicle trajectory generation [8, 319, 320].
For example, the authors in Ref. [8] proposed RC-GAN, a recurrent conditional GAN model that in-
corporated LSTM in both the generator and discriminator, enabling the generation of sensor errors

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:22 H. Yan and Y. Li

with long-term temporal correlations. Empirical evaluation on real road data demonstrated that
RC-GAN produced significantly more realistic sensor error simulations. Zhong et al. [319] intro-
duced a scene-level conditional diffusion model that incorporated a spatio-temporal Transformer
architecture to produce realistic and controllable traffic data. A key innovation was incorporating
language instructions through a LLM, which translated user queries into loss functions that guide
the diffusion model. Evaluations demonstrated that this approach not only produced realistic traf-
fic patterns but also effectively adhered to user constraints, outperforming existing methods in
both realism and controllability. Xu et al. [261] designed a generative AI-based autonomous driv-
ing architecture that utilizes large text-to-image models based on DPM to generate conditioned
traffic and driving data in simulations. The approach improved simulation fidelity, providing a
robust tool for testing and validating autonomous driving systems.
Human trajectory generation. Regarding human trajectory generation, numerous ap-
proaches have also been proposed to solve this problem [22, 138, 159, 193, 230, 253]. Xiong et al.
[253] proposed a semantic-guiding GAN model that integrates semantic knowledge to generate
realistic human trajectories. By employing an attention-based generator and a CNN-based dis-
criminator, this approach effectively captured both the sequential and spatial dependencies of
human movements. Experiments validated its utility in simulating epidemic spread scenarios,
achieving high correlation with real data. STAGE [22] employed a multi-task GAN with spa-
tiotemporal knowledge, integrating a multi-stage generator and spatial consistency loss to produce
privacy-preserving synthetic trajectory. Experimental results indicated that STAGE outperformed
baseline models in data distribution alignment, making it a robust tool for applications requir-
ing high data fidelity and privacy protection. G-GAIL [138] integrated game theory with GAIL
to model pedestrian-vehicle interactions as a Markov decision process, enabling realistic simu-
lation of pedestrian behaviors and enhancing autonomous driving scenario testing. The model
demonstrated superior performance in replicating pedestrian behaviors under various conditions,
enhancing the reliability of autonomous driving scenario testing. Rempe et al. [193] introduced
a controlled pedestrian trajectory generation approach using guided diffusion modeling and a
physics-based humanoid controller, generating realistic movements based on user-defined goals
and environmental constraints, thus improving the realism and flexibility of simulations.
Anomaly data generation. Anomaly data such as traffic accident and crash data is essential
in autonomous driving for safety assessment and scenario testing. However, this type of traffic
data is hard to collect. For example, as illustrated in Ref. [19], the ratio of crash data to non-crash
data is 1 to 11,000, indicating a significant imbalance. Several works use generative AI techniques
to address this problem [30, 35, 96, 99, 286]. For instance, Islam et al. [99] employed VAE to en-
code imbalanced crash and non-crash events into a latent space for generating realistic crash data.
This approach not only achieved statistical similarity with real crash data but also significantly
improved prediction model performance. In Ref. [96], a text-to-traffic GAN model was proposed,
which integrates traffic data with semantic information collected from social media. By incorpo-
rating a global-local loss to bridge the modality gap between textual and traffic data, their model
achieved superior authenticity in generated traffic scenarios. Zarei et al. [286] developed a condi-
tional GAN-based model that utilizes crash counts as conditions to generate realistic crash data.
Their method demonstrated superior performance in hotspot identification, prediction accuracy,
and dispersion parameter estimation, especially in low-dispersion scenarios.

5.3 Traffic Flow Generation

Traffic flow generation refers to generate vehicular traffic volumes or crowd flow within a spe-
cific area or between regions in a transportation network. Numerous works adopt generative AI
techniques to address this task [34, 133, 196, 197, 323]. Specifically, Li et al. [133] presented an

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:23

attentive dual-head spatial-temporal GAN model for generating crowd flow data. This model in-
corporated an attentive mechanism for tracking temporal changes and a self-attention network
to capture intricate spatiotemporal crowd flow patterns, and it adopted a dual-head discriminator
that implements a training strategy with two distinct objectives to mitigate the adverse effects of
rapid overfitting. Zhou et al. [323] developed a knowledge-aware spatio-temporal diffusion model
that utilizes an urban knowledge graph to generate dynamic urban flow data. By integrating re-
gion features and environmental factors, this model accurately generated urban flows for regions
without historical data. In Ref. [196], the authors employed a graph denoising diffusion technique
to generate urban OD flow, effectively capturing the interconnections of nodes and edges within
large urban networks. By decomposing the generative process into network topology and edge
weight generation, their model achieved superior realism in OD matrices across multiple cities.

5.4 Discussion
The literature review highlights several key lessons regarding the application of generative AI
techniques in traffic simulation. Firstly, generative models in Refs. [99, 286] effectively address data
imbalance by synthesizing realistic crash and rare event data. Secondly, integrating multi-modal
and semantic information, as seen in T2 GAN [96], enriches traffic scenario generation by combin-
ing traffic data with social media insights, resulting in more authentic and contextually relevant
simulations. Moreover, privacy-preserving techniques like STAGE [22] ensure data fidelity while
protecting individual privacy, making synthetic data suitable for sensitive applications. However,
these advanced models often face challenges related to computational complexity and scalability,
highlighting the need for further optimization to enable real-time and large-scale implementations.

6 Generative AI for Traffic Decision-making

Traffic decision-making involves a series of decision-making processes made by vehicles or other
road users based on dynamic traffic environment and specific task requirements. Among these
processes, route planning, traffic signal control (TSC) and vehicle control are three fundamental
aspects. Route planning generally refers to generating an optimal path from a designated starting
point to a destination, considering factors such as departure time, road conditions, and traffic
patterns. TSC plays a crucial role in coordinating the movement of vehicles and pedestrians at
intersections, thereby reducing traffic congestion. Vehicle control involves managing the vehicle’s
operation, including actions such as acceleration, steering, and braking, to navigate safely along
driving route.
Ensuring that route planning, TSC, and vehicle control collaboratively make safe and accurate
decisions in complex environments is a formidable challenge. One significant challenge in traffic
decision-making lies in the unpredictable dynamics of traffic environments. These environments
are constantly changing due to various factors, including road conditions, unexpected moves by
human drivers, and sudden decisions by pedestrians and cyclists. For example, a driver might
suddenly brake or change lanes without warning, or pedestrians might cross the road unexpect-
edly. These uncertainties compel autonomous systems to be perpetually alert and adaptive, which
becomes computationally intensive and requires robust prediction algorithms. Another critical
challenge involves rare but critical events that can pose significant risks to safety, such as a child
running onto the road or a mechanical failure causing loss of control. Collecting real-world data
on these rare events is difficult but crucial for creating a comprehensive training dataset. Without
exposure to these scenarios, autonomous systems may not know how to handle them properly,
potentially leading to safety risks. Additionally, the limitations of sensors (e.g., LiDAR, RADAR,
and cameras) present another obstacle. These sensors often have restrictions regarding their oper-
ational range and resolution, as well as their susceptibility to adverse weather conditions, including

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:24 H. Yan and Y. Li

Fig. 7. The relation between different topics in generative AI for traffic decision-making research.

fog, rain, and snow. Such limitations can introduce significant uncertainties, resulting in subopti-
mal decision-making and compromising both safety and efficiency within traffic operations.
Traditional decision-making methods, such as rule-based methods [51, 170] and game theory-
based methods [87, 233, 314], often have limited effectiveness in dealing with complex traffic
situations, while reinforcement learning (RL), as a flexible, efficient, and powerful method,
has gained ground in traffic decision-making. However, the training complexity associated with
handling multiple tasks limits its widespread applicability. Generative AI technology offers
promising solutions to these challenges by leveraging its strengths in data generation, reasoning,
and sequence modeling. For instance, its sequence modeling capabilities help in understanding
and predicting complex driving patterns over time. Furthermore, generative AI’s strong analytical
and reasoning abilities allow for more effective interpretation of external traffic conditions and
instructions, thereby providing important guidance for accurate decision-making. To mitigate
the impact of rare events, generative AI can generate realistic training scenarios mimicking the
unpredictability of real-world driving, reducing the need for extensive road testing. By producing
rare and critical traffic events, generative AI ensures autonomous systems are robust against
challenging situations. Furthermore, generative AI can enhance sensor data, enabling vehicles to
handle challenging environmental conditions with greater adaptability and accuracy.
In the following subsections, we conduct a comprehensive literature review and explore how
generative AI can enhance decision-making processes within traffic systems, specifically focusing
on route planning, traffic signal control, and vehicle control. Their relationship is illustrated in
Figure 7.

6.1 Route Planning

Generative approaches have become integral to route planning tasks within ITS, leveraging their
robust data generation and powerful learning capabilities to incorporate global geographical
knowledge.
One category of approaches utilizes GANs to produce realistic and efficient routes [40, 71, 231].
For example, ProgRPGAN [71] employed a progressive route planning GAN to create realistic
paths across multi-level map resolutions. Similarly, Choi et al. [40] developed a GAN-based frame-
work that integrates feature extraction and path generation networks to produce multiple plausible
routes from egocentric images without relying on high-definition maps.
Another category of approaches integrates RL with generative models to achieve expert-level
path planning. MixGAIL [271] combined transition-aware adversarial imitation learning and GAIL
techniques within a mixed policy gradient actor-critic framework. Building on this, a VAE-GAN
framework for path optimization [236] combined VAEs with GANs to address the vehicle routing
problem on graphs. This approach utilized GNNs to model the complex relationships within road
networks and employed variational reasoning to identify optimal nodes by dividing the graph into

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:25

subgraphs centered around a root node. By integrating RL, the framework achieved end-to-end
optimization, enhancing both learning efficiency and generalization capabilities.
In addition, the integration of LLMs represents a great advancement in providing explainable
and adaptive path planning. DynamicRouteGPT [324] utilized causal inference and pretrained
language models to offer explainable and adaptive real-time path planning in complex traffic sce-
narios. Addressing the limitations of traditional algorithms in large-scale environments, LLM-A*
[171] proposed a novel path planning method that synergistically combines the precise pathfind-
ing capabilities of A* with the global reasoning abilities of LLMs. This hybrid approach aims to
enhance pathfinding efficiency in terms of time and space complexity while maintaining path
validity.

6.2 Traffic Signal Control

Generative methods greatly improve the effectiveness of TSC by improving data integrity, scala-
bility, and decision-making efficiency. The existing literature on generative methods for TSC can
be divided into two categories: single-intersection and multi-intersection control approaches.
Single-intersection control. Generative approaches improve single-intersection traffic signals
by recovering missing data. For instance, DiffLight [27] addressed missing data in TSC by combin-
ing data imputation with decision-making, using a diffusion model to ensure robust control even
with incomplete data. Other research focuses focus on the integration of LLMs with RL [181, 234].
The authors in Ref. [234] introduced a hybrid framework that integrates LLMs with perception and
decision-making tools, enabling the system to adapt to diverse traffic conditions without additional
training. Pang et al. [181] further advanced this integration by employing LLMs to evaluate and
adjust RL decisions under degraded communication and rare events. These integrative methods
demonstrate substantial advantages in adaptability and resilience, effectively addressing the limi-
tations inherent in traditional RL-based systems when confronted with dynamic and unpredictable
traffic environments.
Multi-intersection coordination. This category focuses on data recovery and the coordina-
tion of traffic signals across multiple intersections. A decentralized adaptive TSC framework [242]
employed a GAN-based algorithm for efficient traffic data recovery. This approach significantly
reduced vehicle travel time and maintained high throughput by enabling collaboration with neigh-
boring intersections without heavy data exchange. Similarly, Xu et al. [258] leveraged a WGAN
to estimate agent state spaces for multi-intersection settings, which effectively handles missing
data and enhances system robustness. When integrated with a graph-based deep RL model, this
approach outperformed baseline models in handling complex interactions and incomplete data in
multi-intersection environments.

6.3 Vehicle Control

Vehicle control primarily involves exploring decision-making approaches in autonomous driving
scenarios, including individual vehicle control and connected vehicle control strategies.
Individual vehicle. This subsection discusses methods to enhance decision-making for indi-
vidual vehicles using LLM-assisted approaches and other generative techniques.
LLM-assisted approaches. The integration of LLMs into autonomous vehicle systems has
emerged as a transformative approach to enhancing decision-making capabilities in data analysis,
reasoning, and sequential decision modeling. By leveraging LLMs, these systems can effectively
interpret and process multi-modal external data, perform complex reasoning akin to human
commonsense, and generate coherent decision sequences that enhance navigation and safety. For
instance, DriveGPT4 [263] and TrafficGPT [298] illustrated how LLMs can process multi-frame
video inputs and traffic data to produce vehicle controls and decompose complex tasks, respec-

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:26 H. Yan and Y. Li

tively. This capability significantly enhances the system’s ability to manage intricate driving
scenarios. Additionally, frameworks proposed in works [46] and [45] incorporated environmental
information and driver commands, enabling human-like reasoning and real-time optimization
of driving maneuvers. These frameworks also personalized user experiences and foster trust
through explainable decisions. Furthermore, cognitive reasoning pathways and translation
algorithms, as employed in the literature [204], ensured seamless integration of high-level LLM
decisions with low-level vehicle controls, thereby enhancing adaptability and reliability. Moreover,
models such as the multi-task decision-making GPT [154] and the sequence modeling approach
within a constrained Markov decision process [153] explored the capability of LLMs to handle
multiple driving tasks simultaneously and improve exploration and safety through advanced
techniques like entropy regularization. The use of graph-of-thought structures in explainable
decision-making [39] highlighted the importance of transparency and trust by generating natural
language rationales for driving choices. In summary, this integration not only addresses critical
challenges such as safety assurance, generalization to rare events, and interpretability but also
significantly improves the safety, efficiency, and reliability of autonomous driving, paving the
way for more intelligent and trustworthy transportation solutions.
Other generative techniques. Besides LLM-aided approaches, other generative techniques pro-
vide great advancements in autonomous vehicle decision-making by enhancing robustness and
safety in complex driving scenarios [43, 155, 269]. For example, the authors in Ref. [43] propose
a multi-level GAIL framework that uses a GAN to generate Bird’s-Eye View representations
and a GAIL module to learn vehicle control from these representations, effectively decoupling
representation learning from the driving task to achieve robust autonomous navigation. DDM-Lag
[155] used a diffusion-based generative process to model autonomous vehicle decision-making,
incorporated Lagrangian-based safety enhancements, and integrated hybrid policy updates with
safety-constrained optimization to enhance decision performance in complex driving scenar-
ios. This approach addresses the critical need for safety assurance in AVs by mitigating risks
associated with decision-making under uncertainty, ensuring reliable performance in intricate
driving situations.
Connected vehicles. Recent advancements in connected vehicle control demonstrate the
significant benefits of integrating LLMs into autonomous driving frameworks. CoMAL [273]
leveraged LLMs to facilitate collaboration among autonomous vehicles, enhancing traffic flow and
reducing congestion. A key component of CoMAL is the reason engine, which processes scenario
descriptions and system prompts as inputs. Using a hierarchical chain-of-thought prompt, the
LLM generated driving strategies based on the intelligent driver model, allowing each vehicle
to respond effectively to dynamic situations. This approach not only outperformed traditional
RL methods on the Flow benchmark but also highlighted the strong cooperative capabilities of
LLM-driven agents. Similarly, CoDrivingLLM [63] employed LLMs within a reasoning module for
decision-making, incorporating state perception, intent sharing, negotiation, and chain-of-thought
processing to produce final actions. Experimental results showed that CoDrivingLLM surpasses
rule-based, optimization-based, and machine learning methods, achieving higher success rates,
effective conflict resolution, and balanced efficiency in complex multi-vehicle traffic scenarios.

6.4 Discussion
While existing literature has made significant progress in using generative AI to enhance the ef-
fectiveness of traffic decision-making, there are still areas for improvement in capturing the un-
predictable dynamics of traffic environments, addressing rare but critical events, and overcoming
the limitations of sensors. For instance, current LLM-aided methods such as DriveGPT4 [263] and
TrafficGPT [298], although capable of processing multi-frame video and traffic data, still require

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:27

enhancements in decision robustness and real-time responsiveness in unknown scenarios. Addi-

tionally, while methods like DiffusionES [269] attempt to augment the flexibility and diversity of
autonomous vehicle decision-making strategies, the diversity and authenticity of generating ex-
tremely rare events remain to be validated. Furthermore, the literature provides limited discussion
on enhancing perception and decision-making capabilities under adverse environmental condi-
tions. Addressing it will be a crucial area for future research.

7 Open Challenges and Future Directions

While generative AI methods have demonstrated much success in the realm of ITS, several chal-
lenges must be addressed. In this section, we will discuss the primary open challenges, followed
by an introduction to some potential research directions.

7.1 Open Challenges

In this subsection, we will discuss several open challenges for generative AI methods in ITS.
Multi-modal traffic data. Multi-modal traffic data combines various sources, such as road-
side camera videos, traffic signals, weather conditions, incident reports, and vehicle GPS trajecto-
ries, providing a comprehensive view of traffic conditions. However, using generative AI to solve
traffic-related problems with multi-modal data presents challenges. Firstly, the correlations be-
tween multi-modal data are more complex compared with single-modal data, as the character-
istics and distributions of images, text, and other formats differ significantly. Secondly, accurate
alignment between modalities, such as matching traffic camera images with sensor or weather
data, is critical. Additionally, training models with multi-modal data can be unstable, especially
when maximizing cross-modal information. Therefore, how to effectively use generative AI tech-
nology to tackle transportation challenges using multi-modal data remains challenging in the field
of transportation.
Complex spatio-temporal correlations. Traffic involves complex relationships across time,
space, and spatiotemporal interactions. In the temporal dimension, traffic data is influenced by
factors like weekdays, weekends, and holidays, and current traffic flow can be impacted by pat-
terns from the preceding hours. In the spatial dimension, traffic conditions at different locations
are interrelated, requiring analysis of correlations between traffic flow and congestion across loca-
tions. Additionally, traffic flow at one location may be affected by nearby traffic in the recent past.
Previous research has employed RNNs, CNNs, and GNNs to model spatiotemporal relationships
in traffic, showing promising results. However, most approaches focus on local or micro-scale
spatiotemporal modeling and fail to capture global or macro-scale spatiotemporal patterns. Fur-
thermore, spatiotemporal modeling has yet to fully utilize the powerful inference capabilities of
generative AI. The main challenge lies in effectively modeling the intricate spatiotemporal corre-
lations within traffic data to address traffic-related issues.
Sparse or missing traffic data. Traffic data sparsity or gaps can arise due to equipment failures,
maintenance, network issues, or the costs and privacy concerns of data collection. Such data gaps
can lead to inaccurate traffic predictions and incomplete insights for traffic management. Genera-
tive AI, with its powerful data generation capabilities, has been used to address this issue through
data imputation, traffic estimation, and scene generation. However, there are still shortcomings.
First, due to the spatiotemporal complexity of traffic data, generative AI models may produce data
that appears realistic but lacks spatial or temporal consistency. Second, when generating data, the
generative model may tend to generate common patterns in training data while neglecting less
common cases. This may lead to a lack of diversity in the generated data and inability to cover all
the changes in traffic data. Third, different levels of data sparsity can significantly affect generation
quality, requiring models to adapt flexibly. Therefore, overcoming the challenges of data sparsity

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:28 H. Yan and Y. Li

and absence in addressing traffic issues with generative AI technology remains an important prob-
lem that needs further research.
Adversarial attacks. In intelligent transportation, generative AI technology offers cutting-
edge solutions for traffic management and optimization. However, its widespread adoption also
introduces new security risks, particularly from adversarial attacks [69]. One major concern is
the backdoor attack, where malicious data injected during training can cause models to produce
harmful outputs under specific conditions. In complex intelligent transportation networks, a sin-
gle misjudgment malicious data injected during training can cause models to produce harmful
outputs under specific conditions. For instance, an intelligent transportation monitoring system
compromised by a backdoor attack may erroneously classify a safe vehicle as a potential threat,
or an autonomous vehicle may make incorrect driving decisions as a result. Therefore, the criti-
cal challenge lies in accurately identifying potential risk points and devising effective methods to
ensure the efficient and secure operation of transportation systems.
Model interpretability. Model interpretability is crucial in ITS. For example, in autonomous
driving scenarios, predicting the behavior of other road users, such as pedestrians, cyclists, and
vehicles, is crucial for safe driving. This requires the model to explain why a particular behav-
ior is considered the most likely, such as why a pedestrian may cross the road. While generative
AI models excel in processing traffic data and optimizing transportation systems, they are often
seen as black boxes, which makes it difficult to understand their decision-making. Designing in-
terpretable generative AI models is challenging due to the complexity of traffic systems, involving
numerous variables and uncertainties like road conditions, traffic flow, and pedestrian behavior. In
autonomous driving, autonomous driving decisions involve multiple levels, including perception,
planning, and control, and designing an interpretable model requires being able to explain the
decision-making processes at these different levels, which are often very complex.
Real-time requirements. In ITS, timely and effective response measures are crucial, especially
in rapidly changing situations or unforeseen events. Traffic models must have real-time percep-
tion capabilities to detect vehicles, pedestrians, and obstacles while making immdiate decisions to
ensure safety. Additionally, they must predict changes in traffic flow to take proactive measures
against congestion and accidents. However, achieving this high level of responsiveness poses some
challenges in generative AI-based solutions. First, real-time data may be affected by noise and in-
completeness, so models must have the ability to handle imperfect data to accurately perceive
the traffic environment. Second, certain generative AI methods may be very complex and require
more time for reasoning. In real-time applications, it is necessary to balance the complexity and
real-time performance of the model. Third, real-time applications should be equipped to handle
exceptional circumstances like sensor failures or sudden events. Generative AI methods need to
be capable of identifying such anomalies and executing appropriate actions to ensure both safety
and reliability.
High-perfomance computing. In large metropolis, high population density results in diverse
activities, complex travel behaviors, and numerous transportation routes, all of which present sig-
nificant challenges for evaluating and optimizing traffic management in ITS [1]. Further, leveraging
generative AI for large-scale traffic simulations and optimizations involves running multiple sim-
ulations simultaneously, performing intensive analyses to find optimal solutions, and processing
large volumes of data. High-Performance Computing (HPC) offers opportunities to provide
the required computational power to meet these intensive demands. However, applying HPC to
support generative AI in ITS presents several challenges. As the massive amounts of data generated
from numerous sources in ITS applications, highly efficient data handling and parallel processing
are crucial for enabling timely decision-making. Another key challenge is ensuring scalability and
efficiency in real-time applications. ITS environments are highly dynamic, with traffic conditions

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:29

changing rapidly due to factors such as accidents, weather changes, or special events. The HPC
framework must be able to scale computational resources to handle sudden spikes in data volume
or computational demand, ensuring the system remains responsive under varying conditions.

7.2 Potential Research Directions

Large-scale language models designed using generative AI technology have achieved significant
success in fields such as NLP and image processing, which also provide crucial opportunities and
research directions for addressing traffic-related challenges. In this subsection, we will introduce
several important research directions from three distinct perspectives concerning large-scale traf-
fic models.
Integrating multi-modal data for large-scale traffic models. Traffic data usually come from mul-
tiple different sources, including images and video data from traffic cameras, as well as GPS tra-
jectory data from vehicles, covering a variety of different data types. These data offer multiple
perspectives on the spatiotemporal dynamics of traffic conditions, facilitating a deeper compre-
hension of traffic operation’s underlying mechanisms and patterns. Therefore, these data serve as
an important foundation for training large-scale models in the field of transportation. Efficiently
using these data for model training is crucial and also an important research direction. In addition
to aligning the data in the spatiotemporal dimension, it is necessary to unify their representation,
mapping different data modalities to the same data space. This requires designing appropriate
mapping methods to ensure that the mapped data can effectively capture spatial and temporal
dependencies across different data modalities.
Designing and training well-performed large-scale traffic models. Traditional transportation mod-
els are typically developed separately for specific tasks, requiring significant time and resources
to adapt to new requirements. These models also tend to be overly specialized, making cross-task
transfer learning difficult and limiting scalability. In contrast, large-scale language models based on
generative AI technology have higher generalizability and scalability, with the potential for trans-
fer learning across various traffic tasks. For instance, in autonomous driving, which encompasses
continuous tasks such as perception, prediction, and decision-making, a large-scale model could
handle all tasks simultaneously by modeling shared features like spatiotemporal characteristics
while considering each task’s unique aspects. Recently, GPT-ST [141] gives a preliminary attempt
in designing generative spatio-temporal pre-training framework. Another promising direction is
developing large-scale traffic models driven by both data and knowledge, combining diverse traffic
data with reliable physical knowledge to effectively model traffic patterns.s
Traffic planning and decision-making based on large-scale language models. Traffic planning and
decision-making are important research problems in the transportation domain. In real-world sce-
narios, traffic conditions constantly change due to congestion, accidents, or adverse weather, im-
pacting the effectiveness of planning and decision-making. Therefore, it is essential to develop
real-time traffic planning and decision-making algorithms based on large-scale language models.
This entails exploring how to combine the real-time traffic perception capabilities of generative AI
technology with the decision-making capabilities of RL. Such an approach holds the potential to
yield more efficient and adaptable solutions to tackle the intricacies and uncertainties inherent in
traffic management. As an illustrative example, consider driving decision-making in autonomous
vehicles. Vehicles need to make decisions such as overtaking, decelerating, and stopping based on
perceived data and environmental conditions. Large-scale models, equipped with powerful knowl-
edge reasoning and generation capabilities, are expected to achieve precise traffic decisions in
complex traffic environments. Therefore, future research can focus on how to design algorithms
based on large-scale language models to make accurate decisions while ensuring a balance in real-
time performance and safety.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:30 H. Yan and Y. Li

8 Conclusions
Generative AI technology is playing an increasingly influential role in ITS, particularly in per-
ception, prediction, simulation, and decision-making. In this survey, we systematically investigate
key generative AI solutions within ITS, and explore the challenges that remain unresolved in this
field. Our analysis identifies several valuable guidelines. First, designing specialized HPC allows for
large-scale traffic simulations and optimizations with massive multi-modal data. Second, ensuring
data reliability and system security requires addressing data sparsity and safeguarding against ad-
versarial attacks. Third, enhancing real-time performance while improving model interpretability
is crucial to enable prompt responses to dynamic traffic conditions and foster trust through trans-
parent decision-making. Based on these guidelines, we summarize several key trends in generative
AI for ITS. One significant trend is the integration of multi-modal data to enable comprehensive
and large-scale traffic modeling. Another is the focus on designing generalizable models capable
of addressing multiple transportation tasks simultaneously, which enhances the versatility and
robustness of AI solutions. Additionally, there is growing interest in developing real-time traffic
planning and decision-making algorithms, particularly by leveraging the capabilities of large-scale
language models. These trends illustrate the transformative potential of generative AI in reshaping
ITS, paving the way for safer, more efficient, and resilient transportation systems.

References
[1] Aya Aboudina, Islam Kamel, Mohamed Elshenawy, Hossam Abdelgawad, and Baher Abdulhai. 2018. Harnessing the
power of hpc in simulation and optimization of large transportation networks: Spatio-temporal traffic management
in the greater toronto area. IEEE Intelligent Transportation Systems Magazine 10, 1 (2018), 95–106.
[2] Abasi-Amefon O. Affia, Raimundas Matulevičius, and Alexander Nolte. 2019. Security risk management in coopera-
tive intelligent transportation systems: A systematic literature review. In On the Move to Meaningful Internet Systems:
OTM 2019 Conferences: Confederated International Conferences: CoopIS, ODBASE, C&TC 2019, Rhodes, Greece, October
21–25, 2019, Proceedings. Springer, 282–300.
[3] Qazi Ejaz Ali, Naveed Ahmad, Abdul Haseeb Malik, Gauhar Ali, and Waheed Ur Rehman. 2018. Issues, challenges,
and research opportunities in intelligent transport system for security and privacy. Applied Sciences 8, 10 (2018), 1964.
[4] Insha Altaf and Ajay Kaul. 2021. A survey on autonomous vehicles in the field of intelligent transport system. In
Applications of Networks, Sensors and Autonomous Systems Analytics: Proceedings of ICANSAA 2020. Springer, 11–31.
[5] Mohammed Arif Amin, Samah Hadouej, and Tasneem S. J. Darwish. 2019. Big data role in improving intelligent
transportation systems safety: A survey. In Advances in Internet, Data and Web Technologies: The 7th International
Conference on Emerging Internet, Data and Web Technologies (EIDWT’19). Springer, 187–199.
[6] Sheng-hai An, Byung-Hyug Lee, and Dong-Ryeol Shin. 2011. A survey of intelligent transportation systems. In 2011
Third International Conference on Computational Intelligence, Communication Systems and Networks. IEEE, 332–337.
[7] Sesham Anand, P. Padmanabham, A. Govardhan, and Rajesh H. Kulkarni. 2018. An extensive review on data min-
ing methods and clustering models for intelligent transportation system. Journal of Intelligent Systems 27, 2 (2018),
263–273.
[8] Henrik Arnelid, Edvin Listo Zec, and Nasser Mohammadiha. 2019. Recurrent conditional generative adversarial net-
works for autonomous driving sensor modelling. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC’19).
IEEE, 1613–1618.
[9] Nazia Aslam and Maheshkumar H. Kolekar. 2023. A-VAE: Attention based variational autoencoder for traffic video
anomaly detection. In 2023 IEEE 8th International Conference for Convergence in Technology (I2CT’23). IEEE, 1–7.
[10] Han Bao, Xun Zhou, Yiqun Xie, Yingxue Zhang, and Yanhua Li. 2022. COVID-GAN+: Estimating human mobility
responses to COVID-19 through spatio-temporal generative adversarial networks with enhanced features. ACM
Transactions on Intelligent Systems and Technology (TIST) 13, 2 (2022), 1–23.
[11] Han Bao, Xun Zhou, Yingxue Zhang, Yanhua Li, and Yiqun Xie. 2020. Covid-gan: Estimating human mobility re-
sponses to covid-19 pandemic through spatio-temporal conditional generative adversarial networks. In Proceedings
of the 28th International Conference on Advances in Geographic Information Systems. 273–282.
[12] Naren Bao, Alexander Carballo, and Takeda Kazuya. 2021. Prediction of personalized driving behaviors via driver-
adaptive deep generative models. In 2021 IEEE Intelligent Vehicles Symposium (IV’21). IEEE, 616–621.
[13] Sarah Baras, Iman Saeed, Hadeel A. Tabaza, and Mourad Elhadef. 2018. VANETs-based intelligent transportation
systems: An overview. Advances in Computer Science and Ubiquitous Computing: CSA-CUTE 17 474 (2018), 265–273.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:31

[14] Elyes Ben Hamida, Hassan Noura, and Wassim Znaidi. 2015. Security of cooperative intelligent transport systems:
Standards, threats analysis and cryptographic countermeasures. Electronics 4, 3 (2015), 380–423.
[15] Kartik Krishna Bhardwaj, Anirudh Khanna, Deepak Kumar Sharma, and Anshuman Chhabra. 2019. Designing
energy-efficient IoT-based intelligent transport system: need, architecture, characteristics, challenges, and applica-
tions. Energy Conservation for IoT Devices: Concepts, Paradigms and Solutions 206 (2019), 209–233.
[16] Alden Boby, Dane Brown, and James Connan. 2023. Iterative refinement versus generative adversarial networks for
super-resolution towards licence plate detection. In Inventive Systems and Control: Proceedings of ICISC 2023. Springer,
349–362.
[17] Guillem Boquet, Antoni Morell, Javier Serrano, and Jose Lopez Vicario. 2020. A variational autoencoder solution
for road traffic forecasting systems: Missing data imputation, dimension reduction, model selection and anomaly
detection. Transportation Research Part C: Emerging Technologies 115 (2020), 102622.
[18] Guillem Boquet, Jose Lopez Vicario, Antoni Morell, and Javier Serrano. 2019. Missing data in traffic estimation: A
variational autoencoder imputation method. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP’19). IEEE, 2882–2886.
[19] Qing Cai, Mohamed Abdel-Aty, Jinghui Yuan, Jaeyoung Lee, and Yina Wu. 2020. Real-time crash prediction on ex-
pressways using deep generative models. Transportation Research Part C: Emerging Technologies 117 (2020), 102697.
[20] Fernando Camacho, César Cárdenas, and David Muñoz. 2018. Emerging technologies and research challenges for
intelligent transportation systems: 5G, HetNets, and SDN. International Journal on Interactive Design and Manufac-
turing (IJIDeM) 12 (2018), 327–335.
[21] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S. Yu, and Lichao Sun. 2023. A comprehensive survey
of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv:2303.04226. Retrieved from
https://arxiv.org/abs/2303.04226
[22] Zhongcai Cao, Kang Liu, Xin Jin, Li Ning, Ling Yin, and Feng Lu. 2024. STAGE: A spatiotemporal-knowledge en-
hanced multi-task generative adversarial network (GAN) for trajectory generation. International Journal of Geo-
graphical Information Science 39, 5 (2024), 1100–1127.
[23] H. Varun Chand and J. Karthikeyan. 2018. Survey on the role of IoT in intelligent transportation system. Indonesian
Journal of Electrical Engineering and Computer Science 11, 3 (2018), 936–941.
[24] Chuan Chen, Zhenpeng Wu, Yanyi Lai, Wenlin Ou, Tianchi Liao, and Zibin Zheng. 2023. Challenges and reme-
dies to privacy and security in AIGC: Exploring the potential of privacy computing, blockchain, and beyond.
arXiv:2306.00419. Retrieved from https://arxiv.org/abs/2306.00419
[25] Di Chen, Meixin Zhu, Hao Yang, Xuesong Wang, and Yinhai Wang. 2024. Data-driven traffic simulation: A compre-
hensive review. IEEE Transactions on Intelligent Vehicles 9, 4 (2024), 4730–4748.
[26] Haonan Chen, Tianchen Ji, Shuijing Liu, and Katherine Driggs-Campbell. 2022. Combining model-based controllers
and generative adversarial imitation learning for traffic simulation. In 2022 IEEE 25th International Conference on
Intelligent Transportation Systems (ITSC’22). IEEE, 1698–1704.
[27] Hanyang Chen, Yang Jiang, Shengnan Guo, Xiaowei Mao, Youfang Lin, and Huaiyu Wan. 2024. DiffLight: A partial
rewards conditioned diffusion model for traffic signal control with missing data. In Advances in Neural Information
Processing Systems.
[28] Jiayuan Chen, Shuo Zhang, Xiaofei Chen, Qiao Jiang, Hejiao Huang, and Chonglin Gu. 2021. Learning traffic as
videos: A spatio-temporal VAE approach for traffic data imputation. In International Conference on Artificial Neural
Networks. Springer, 615–627.
[29] Kehua Chen, Xianda Chen, Zihan Yu, Meixin Zhu, and Hai Yang. 2023. EquiDiff: A conditional equivariant diffusion
model for trajectory prediction. 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC).
IEEE, 746–751.
[30] Mu-Yen Chen, Hsiu-Sen Chiang, and Wei-Kai Huang. 2022. Efficient generative adversarial networks for imbalanced
traffic collision datasets. IEEE Transactions on Intelligent Transportation Systems 23, 10 (2022), 19864–19873.
[31] Qi Chen, Wei Wang, Kaizhu Huang, Suparna De, and Frans Coenen. 2021. Multi-modal generative adversarial net-
works for traffic event detection in smart cities. Expert Systems with Applications 177 (2021), 114939.
[32] Xinyu Chen, Jiajie Xu, Rui Zhou, Wei Chen, Junhua Fang, and Chengfei Liu. 2021. Trajvae: A variational autoencoder
model for trajectory generation. Neurocomputing 428 (2021), 332–339.
[33] Yuanyuan Chen, Yisheng Lv, and Fei-Yue Wang. 2019. Traffic flow imputation using parallel data and generative
adversarial networks. IEEE Transactions on Intelligent Transportation Systems 21, 4 (2019), 1624–1630.
[34] Yuanyuan Chen, Yisheng Lv, and Fenghua Zhu. 2021. Traffic flow synthesis using generative adversarial networks via
semantic latent codes manipulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC’21).
IEEE, 1451–1456.
[35] Zhijun Chen, Jingming Zhang, Yishi Zhang, and Zihao Huang. 2021. Traffic accident data generation based on im-
proved generative adversarial networks. Sensors 21, 17 (2021), 5767.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:32 H. Yan and Y. Li

[36] Hao Cheng, Li Feng, Hailong Liu, Takatsugu Hirayama, Hiroshi Murase, and Monika Sester. 2021. Interaction de-
tection between vehicles and vulnerable road users: A deep generative approach with attention. arXiv:2105.03891.
Retrieved from https://arxiv.org/abs/2105.03891
[37] Hao Cheng, WLMY Yang, M. Sester, and B. Rosenhahn. 2020. Context conditional variational autoencoder for pre-
dicting multi-path trajectories in mixed traffic. arXiv:2002.05966. Retrieved from https://arxiv.org/abs/2002.05966
[38] Rishu Chhabra, Seema Verma, and C. Rama Krishna. 2017. A survey on driver behavior detection techniques for intel-
ligent transportation systems. In 2017 7th International Conference on Cloud Computing, Data Science & Engineering-
Confluence. IEEE, 36–41.
[39] Fangyuan Chi, Yixiao Wang, Panos Nasiopoulos, and Victor C. M. Leung. 2024. Multi-modal GPT-4 aided action
planning and reasoning for self-driving vehicles. In ICASSP 2024-2024 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP’24). IEEE, 7325–7329.
[40] Dooseop Choi, Seung-Jun Han, Kyoung-Wook Min, and Jeongdan Choi. 2022. PathGAN: Local path planning with
attentive generative adversarial networks. ETRI Journal 44, 6 (2022), 1004–1019.
[41] Junsung Choi, Vuk Marojevic, Carl B. Dietrich, Jeffrey H. Reed, and Seungyoung Ahn. 2020. Survey of spectrum
regulation for intelligent transportation systems. IEEE Access 8 (2020), 140145–140160.
[42] Seongjin Choi, Jiwon Kim, and Hwasoo Yeo. 2021. TrajGAIL: Generating urban vehicle trajectories using generative
adversarial imitation learning. Transportation Research Part C: Emerging Technologies 128 (2021), 103091.
[43] Gustavo Claudio Karl Couto and Eric Aislan Antonelo. 2024. Hierarchical generative adversarial imitation learning
with mid-level input generation for autonomous driving on urban environments. IEEE Transactions on Intelligent
Vehicles (2024), 1–14.
[44] Christian Creß, Zhenshan Bing, and Alois C. Knoll. 2023. Intelligent transportation systems using roadside infras-
tructure: A literature survey. IEEE Transactions on Intelligent Transportation Systems 25, 7 (2023), 6309–6327.
[45] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. 2024. Drive as you speak: Enabling human-like
interaction with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference
on Applications of Computer Vision. 902–909.
[46] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. 2024. Receive, reason, and react: Drive as you say,
with large language models in autonomous vehicles. IEEE Intelligent Transportation Systems Magazine 16, 4 (2024),
81–94.
[47] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang,
Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo
Cao, Ziran Wang, and Chao Zheng. 2024. A survey on multimodal large language models for autonomous driving.
In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 958–979.
[48] Qiongjie Cui, Huaijiang Sun, Yue Kong, Xiaoqian Zhang, and Yanmeng Li. 2021. Efficient human motion prediction
using temporal convolutional generative adversarial network. Information Sciences 545 (2021), 427–447.
[49] Issam Damaj, Salwa K. Al Khatib, Tarek Naous, Wafic Lawand, Zainab Z. Abdelrazzak, and Hussein T. Mouftah. 2022.
Intelligent transportation systems: A survey on modern hardware devices for the era of machine learning. Journal
of King Saud University-Computer and Information Sciences 34, 8 (2022), 5921–5942.
[50] Debashis Das, Sourav Banerjee, Pushpita Chatterjee, Uttam Ghosh, and Utpal Biswas. 2023. Blockchain for intelligent
transportation systems: Applications, challenges, and opportunities. IEEE Internet of Things Journal 10, 21 (2023),
18961–18970.
[51] Pierre De Beaucorps, Thomas Streubel, Anne Verroust-Blondet, Fawzi Nashashibi, Benazouz Bradai, and Paulo Re-
sende. 2017. Decision-making for automated vehicles at intersections adapting human-like behavior. In 2017 IEEE
Intelligent Vehicles Symposium (IV’17). IEEE, 212–217.
[52] Xiaoheng Deng, Leilei Wang, Jinsong Gui, Ping Jiang, Xuechen Chen, Feng Zeng, and Shaohua Wan. 2023. A review
of 6G autonomous intelligent transportation systems: Mechanisms, applications and challenges. Journal of Systems
Architecture 142 (2023), 102929.
[53] Praveen Devadhas Sujakumari and Paulraj Dassan. 2023. Generative adversarial networks (GAN) and HDFS-based
realtime traffic forecasting system using CCTV surveillance. Symmetry 15, 4 (2023), 779.
[54] S. Devika, Rishi Rakesh Shrivastava, Pratik Narang, Tejasvi Alladi, and F. Richard Yu. 2024. VADGAN: An unsuper-
vised GAN framework for enhanced anomaly detection in connected and autonomous vehicles. IEEE Transactions
on Vehicular Technology 73, 9 (2024), 12458–12467.
[55] Shaohua Ding, Yulong Tian, Fengyuan Xu, Qun Li, and Sheng Zhong. 2019. Trojan attack on deep generative mod-
els in autonomous driving. In Security and Privacy in Communication Networks: 15th EAI International Conference,
SecureComm 2019, Orlando, FL, USA, October 23–25, 2019, Proceedings, Part I 15. Springer, 299–318.
[56] Jiqian Dong, Sikai Chen, and Samuel Labi. 2023. Transfusor: Transformer diffusor for controllable human-like gen-
eration of vehicle lane changing Trajectories. arXiv:2308.14943. Retrieved from https://arxiv.org/abs/2308.14943

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:33

[57] Hongyang Du, Zonghang Li, Dusit Niyato, Jiawen Kang, Zehui Xiong, Dong In Kim, et al. 2023. Enabling AI-
generated content (AIGC) services in wireless edge networks. arXiv:2301.03220. Retrieved from https://arxiv.org/
abs/2301.03220
[58] Yilun Du and Igor Mordatch. 2019. Implicit generation and modeling with energy based models. Advances in Neural
Information Processing Systems 32 (2019), 1–11.
[59] Yujun Du, Jinling Wang, Chris Rizos, and Ahmed El-Mowafy. 2021. Vulnerabilities and integrity of precise point
positioning for intelligent transport systems: Overview and analysis. Satellite Navigation 2, 1 (2021), 1–22.
[60] Nour-Eddin El Faouzi, Henry Leung, and Ajeesh Kurian. 2011. Data fusion in intelligent transportation systems:
Progress and challenges–A survey. Information Fusion 12, 1 (2011), 4–10.
[61] Ayoub Ellahyani, Ilyas El Jaafari, and Said Charfi. 2021. Traffic sign detection for intelligent transportation systems:
A survey. In E3S Web of Conferences, Vol. 229. EDP Sciences, 01006.
[62] Jianwu Fang, Lei-lei Li, Junfei Zhou, Junbin Xiao, Hongkai Yu, Chen Lv, Jianru Xue, and Tat-Seng Chua. 2024. Abduc-
tive ego-view accident video understanding for safe driving perception. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition. 22030–22040.
[63] Shiyu Fang, Jiaqi Liu, Mingyu Ding, Yiming Cui, and Chen Lv. 2025. Towards interactive and learnable cooperative
driving automation: A large language model-driven decision-making framework. IEEE Transactions on Vehicular
Technology.
[64] Danish Fayaz. 2018. Intelligent transport System-A review. [Electronic Resource]. Access mode: https://www.
researchgate.net/publication/329864030_Intelligent_Transport_System-A_Review (access date: 05/30/2025).
[65] Fenling Feng, Jiaqi Zhang, Chengguang Liu, Wan Li, and Qiwei Jiang. 2021. Short-term railway passenger demand
forecast using improved Wasserstein generative adversarial nets and web search terms. IET Intelligent Transport
Systems 15, 3 (2021), 432–445.
[66] Ruyi Feng, Zhibin Li, Bowen Liu, Yan Ding, and Ou Zheng. 2023. TrTr: A versatile pre-trained large traffic model
based on transformer for capturing trajectory diversity in vehicle population. arXiv:2309.12677. Retrieved from https:
//arxiv.org/abs/2309.12677
[67] Tao Feng, Huan Yan, Huandong Wang, Wenzhen Huang, Yuyang Han, Hongsen Liao, Jinghua Hao, and Yong Li.
2023. ILRoute: A graph-based imitation learning method to unveil riders’ routing strategies in food delivery service.
In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4024–4034.
[68] Lino Figueiredo, Isabel Jesus, J. A. Tenreiro Machado, Jose Rui Ferreira, and J. L. Martins De Carvalho. 2001. To-
wards the development of intelligent transportation systems. ITSC 2001. 2001 IEEE Intelligent Transportation Systems.
Proceedings (Cat. No. 01TH8585) (2001), 1206–1211.
[69] Lin Geng Foo, Hossein Rahmani, and Jun Liu. 2023. AIGC for various data modalities: A survey. arXiv:2308.14177.
Retrieved from https://arxiv.org/abs/2308.14177
[70] Marcia R. Friesen and Robert D. McLeod. 2015. Bluetooth in intelligent transportation systems: A survey. Interna-
tional Journal of Intelligent Transportation Systems Research 13 (2015), 143–153.
[71] Tao-yang Fu and Wang-Chien Lee. 2021. ProgRPGAN: Progressive GAN for route planning. In Proceedings of the
27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 393–403.
[72] Tanya Garg and Gurjinder Kaur. 2023. A systematic review on intelligent transport systems. Journal of Computational
and Cognitive Engineering 2, 3 (2023), 175–188.
[73] Tukaram K. Gawali and Shailesh S. Deore. 2024. Dual-discriminator conditional Giza pyramids construction genera-
tive adversarial network based traffic density recognition using road vehicle images. International Journal of Machine
Learning and Cybernetics 15, 3 (2024), 1007–1024.
[74] Benyamin Ghojogh and Ali Ghodsi. 2020. Attention mechanism, transformers, BERT, and GPT: Tutorial and survey.
https://hal.science/hal-04637647v1/document (access date: 05/30/2025).
[75] Ashkan Gholamhosseinian and Jochen Seitz. 2021. Vehicle classification in intelligent transport systems: An
overview, methods and software perspective. IEEE Open Journal of Intelligent Transportation Systems 2 (2021),
173–194.
[76] Arna Ghosh, Biswarup Bhattacharya, and Somnath Basu Roy Chowdhury. 2016. Sad-gan: Synthetic autonomous
driving using generative adversarial networks. arXiv:1611.08788. Retrieved from https://arxiv.org/abs/1611.08788
[77] Raksha Ghosh, R. Pragathi, S. Ullas, and Surekha Borra. 2017. Intelligent transportation systems: A survey. In 2017
International Conference on Circuits, Controls, and Communications (CCUBE’17). IEEE, 160–165.
[78] Taiyuan Gong, Li Zhu, F. Richard Yu, and Tao Tang. 2023. Edge intelligence in intelligent transportation systems: A
survey. IEEE Transactions on Intelligent Transportation Systems 24, 9 (2023), 8919–8944.
[79] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (2020), 139–144.
[80] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
Yoshua Bengio. 2014. Generative adversarial nets. Advances in Neural Information Processing Systems 27 (2014), 1–9.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:34 H. Yan and Y. Li

[81] Artur Grigorev, Khaled Saleh, and Yuming Ou. 2024. IncidentResponseGPT: Generating traffic incident response
plans with generative artificial intelligence. arXiv:2404.18550. Retrieved from https://arxiv.org/abs/2404.18550
[82] Ning Gui, Tianchu Zeng, Jianming Hu, and Yi Zhang. 2022. Visual-angle attention predictor: A multi-agent trajectory
predictor based on variational auto-encoder. In CICTP 2022. 866–877.
[83] Arya Ketabchi Haghighat, Varsha Ravichandra-Mouli, Pranamesh Chakraborty, Yasaman Esfandiari, Saeed Arabi,
and Anuj Sharma. 2020. Applications of deep learning in intelligent transportation systems. Journal of Big Data
Analytics in Transportation 2 (2020), 115–145.
[84] Zahid Halim, Rizwana Kalsoom, and Abdul Rauf Baig. 2016. Profiling drivers based on driver dependent vehicle
driving features. Applied Intelligence 44 (2016), 645–664.
[85] Byung-Gil Han, Jong Taek Lee, Kil-Taek Lim, and Doo-Hyun Choi. 2020. License plate image generation using gener-
ative adversarial networks for end-to-end license plate character recognition from a small set of real images. Applied
Sciences 10, 8 (2020), 2780.
[86] Lingyi Han, Kan Zheng, Long Zhao, Xianbin Wang, and Huimin Wen. 2020. Content-aware traffic data completion
in ITS based on generative adversarial nets. IEEE Transactions on Vehicular Technology 69, 10 (2020), 11950–11962.
[87] Peng Hang, Chen Lv, Yang Xing, Chao Huang, and Zhongxu Hu. 2020. Human-like decision making for autonomous
driving: A noncooperative game theoretic approach. IEEE Transactions on Intelligent Transportation Systems 22, 4
(2020), 2076–2087.
[88] LIN Hao, LI Leixiao, and WANG Hui. 2020. Survey on research and application of support vector machines in intel-
ligent transportation system. Journal of Frontiers of Computer Science & Technology 14, 6 (2020), 901.
[89] Julie Harvey and Sathish Kumar. 2020. A survey of intelligent transportation systems security: challenges and so-
lutions. In 2020 IEEE 6th Intl Conference on Big Data Security on Cloud (BigDataSecurity), IEEE Intl Conference on
High Performance and Smart Computing,(HPSC) and IEEE Intl Conference on Intelligent Data and Security (IDS). IEEE,
263–268.
[90] Muhammad Abul Hassan, Roofia Javed, Fabrizio Granelli, Xin Gen, Muhammad Rizwan, Syed Haider Ali, Hazrat
Junaid, Sana Ullah, et al. 2023. Intelligent transportation systems in smart city: A systematic survey. In 2023 Interna-
tional Conference on Robotics and Automation in Industry (ICRAI’23). IEEE, 1–9.
[91] Md Saif Hassan Onim, Hussain Nyeem, Md Wahiduzzaman Khan Arnob, and Arunima Dey Pooja. 2023. Unleash-
ing the power of generative adversarial networks: A novel machine learning approach for vehicle detection and
localisation in the dark. Cognitive Computation and Systems 5, 3 (2023), 169–180.
[92] Ammar Haydari and Yasin Yılmaz. 2020. Deep reinforcement learning for intelligent transportation systems: A sur-
vey. IEEE Transactions on Intelligent Transportation Systems 23, 1 (2020), 11–32.
[93] Jonathan Ho and Stefano Ermon. 2016. Generative adversarial imitation learning. Advances in Neural Information
Processing Systems 29 (2016), 1–9.
[94] Tongge Huang, Pranamesh Chakraborty, and Anuj Sharma. 2021. Deep convolutional generative adversarial net-
works for traffic data imputation encoding time series as images. International Journal of Transportation Science and
Technology 12, 1 (2021), 1–18.
[95] Yifei Huang and Feng Chen. 2022. Data interpolation of traffic flow algorithm using wavelet transform for traffic
generative modeling. IEEE Journal of Radio Frequency Identification 6 (2022), 739–742.
[96] Guangyu Huo, Yong Zhang, Boyue Wang, Yongli Hu, and Baocai Yin. 2021. Text-to-traffic generative adversarial
network for traffic situation generation. IEEE Transactions on Intelligent Transportation Systems 23, 3 (2021), 2623–
2636.
[97] H Ibrahim, Omar M. Fahmy, and Mustafa A. Elattar. 2022. License plate image analysis empowered by generative
adversarial neural networks (GANs). IEEE Access 10 (2022), 30846–30857.
[98] Davide Imparato, Ahmed El-Mowafy, Chris Rizos, and Jinling Wang. 2018. Vulnerabilities in SBAS and RTK posi-
tioning in intelligent transport systems: An overview. In Proceedings of the International Global Navigation Satellite
System Association IGNSS Symposium.
[99] Zubayer Islam, Mohamed Abdel-Aty, Qing Cai, and Jinghui Yuan. 2021. Crash data augmentation using variational
autoencoder. Accident Analysis & Prevention 151 (2021), 105950.
[100] Boris Ivanovic, Karen Leung, Edward Schmerling, and Marco Pavone. 2020. Multimodal deep generative models for
trajectory prediction: A conditional variational autoencoder approach. IEEE Robotics and Automation Letters 6, 2
(2020), 295–302.
[101] D. N. Jagadish, Arun Chauhan, and Lakshman Mahto. 2021. Autonomous vehicle path prediction using conditional
variational autoencoder networks. In Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer,
129–139.
[102] D. N. Jagadish, Arun Chauhan, and Lakshman Mahto. 2022. Conditional variational autoencoder networks for au-
tonomous vehicle path prediction. Neural Processing Letters 54, 5 (2022), 3965–3978.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:35

[103] Junchen Jin, Dingding Rong, Tong Zhang, Qingyuan Ji, Haifeng Guo, Yisheng Lv, Xiaoliang Ma, and Fei-Yue Wang.
2022. A GAN-based short-term link traffic prediction approach for urban road networks under a parallel learning
framework. IEEE Transactions on Intelligent Transportation Systems 23, 9 (2022), 16185–16196.
[104] Xue-Bo Jin, Wen-Tao Gong, Jian-Lei Kong, Yu-Ting Bai, and Ting-Li Su. 2022. PFVAE: A planar flow-based variational
auto-encoder prediction model for time series data. Mathematics 10, 4 (2022), 610.
[105] Ye Jin, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, and
Jiangtao Gong. 2023. SurrealDriver: Designing generative driver agent simulation framework in urban contexts based
on large language model. arXiv:2309.13193. Retrieved from https://arxiv.org/abs/2309.13193
[106] Levente Juhász, Peter Mooney, Hartwig H. Hochmair, and Boyuan Guan. 2023. ChatGPT as a mapping assistant: A
novel method to enrich maps with generative AI and content derived from street-level photographs. arXiv:2306.03204.
Retrieved from https://arxiv.org/abs/2306.03204
[107] Li-Wei Kang, Chih-Chung Hsu, I-Shan Wang, Ting-Lei Liu, Shih-Yu Chen, and Chuan-Yu Chang. 2020. Vehicle trajec-
tory prediction based on social generative adversarial network for self-driving car applications. In 2020 International
Symposium on Computer, Consumer and Control (IS3C’20). IEEE, 489–492.
[108] Yan Kang, Jinyuan Li, Shin-Jye Lee, and Hao Li. 2020. Generative adversarial network-based regional epitaxial traffic
flow prediction. In Advances in Natural Computation, Fuzzy Systems and Knowledge Discovery: Volume 2. Springer,
804–814.
[109] Zhuangwei Kang, Ayan Mukhopadhyay, Aniruddha Gokhale, Shijie Wen, and Abhishek Dubey. 2022. Traffic anomaly
detection via conditional normalizing flow. In 2022 IEEE 25th International Conference on Intelligent Transportation
Systems (ITSC’22). IEEE, 2563–2570.
[110] Amir Kazemi and Hadi Meidani. 2021. IGANI: Iterative generative adversarial networks for imputation with appli-
cation to traffic data. IEEE Access 9 (2021), 112966–112977.
[111] Alkilane Khaled, Alfateh M. Tag Elsir, and Yanming Shen. 2022. TFGAN: Traffic forecasting using generative adver-
sarial network with multi-graph convolutional network. Knowledge-Based Systems 249 (2022), 108990.
[112] Aidil Redza Khan, Mohd Faizal Jamlos, Nurmadiha Osman, Muhammad Izhar Ishak, Fatimah Dzaharudin, You Kok
Yeow, and Khairil Anuar Khairi. 2022. DSRC technology in Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure
(V2I) IoT system for Intelligent Transportation System (ITS): A review. Recent Trends in Mechatronics Towards Indus-
try 4.0: Selected Articles from iM3F 2020, Malaysia 730 (2022), 97–106.
[113] Dongchan Kim, Hyukju Shon, Nahyun Kweon, Seungwon Choi, Chanuk Yang, and Kunsoo Huh. 2021. Driving style-
based conditional variational autoencoder for prediction of ego vehicle trajectory. IEEE Access 9 (2021), 169348–
169356.
[114] Diederik P. Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv:1312.6114. Retrieved from https:
//arxiv.org/abs/1312.6114
[115] Alireza Koochali, Peter Schichtel, Andreas Dengel, and Sheraz Ahmed. 2019. Probabilistic forecasting of sensory data
with generative adversarial networks–forgan. IEEE Access 7 (2019), 63868–63880.
[116] Robert Krajewski, Tobias Moers, Adrian Meister, and Lutz Eckstein. 2019. BézierVAE: Improved trajectory model-
ing using variational autoencoders for the safety validation of highly automated vehicles. In 2019 IEEE Intelligent
Transportation Systems Conference (ITSC’19). IEEE, 3788–3795.
[117] Robert Krajewski, Tobias Moers, Dominik Nerger, and Lutz Eckstein. 2018. Data-driven maneuver modeling using
generative adversarial networks and variational autoencoders for safety validation of highly automated vehicles. In
2018 21st International Conference on Intelligent Transportation Systems (ITSC’18). IEEE, 2383–2390.
[118] Mahesh Venkata Krishna and Joachim Denzler. 2014. A combination of generative and discriminative models for fast
unsupervised activity recognition from traffic scene videos. In IEEE Winter Conference on Applications of Computer
Vision. IEEE, 640–645.
[119] P. M. Ashok Kumar, D. Kavitha, and S. Arun Kumar. 2020. A hybrid generative-discriminative model for abnormal
event detection in surveillance video scenes. International Journal of Information and Computer Security 12, 2-3 (2020),
253–268.
[120] R. Sampath Kumar, V. P. Krishnamurthy, Venkateswararao Podile, G. Yamini Priyanka, and Vemulapalli Neha. 2023.
Generative adversarial networks to improve the nature of training in autonomous vehicles. In 2023 International
Conference on Disruptive Technologies (ICDT’23). IEEE, 161–164.
[121] Ayyoub Lamssaggad, Nabil Benamar, Abdelhakim Senhaji Hafid, and Mounira Msahli. 2021. A survey on the current
security landscape of intelligent transportation systems. IEEE Access 9 (2021), 9180–9208.
[122] Yann LeCun, Sumit Chopra, Raia Hadsell, M. Ranzato, and Fujie Huang. 2006. A tutorial on energy-based learning.
Predicting Structured Data 1, 0 (2006), 1–59.
[123] Younkwan Lee, JaeWoong Yun, Yoojin Hong, Juhyun Lee, and Moongu Jeon. 2018. Accurate license plate recognition
and super-resolution using a generative adversarial networks on traffic surveillance video. In 2018 IEEE International
Conference on Consumer Electronics-Asia (ICCE-Asia’18). IEEE, 1–4.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:36 H. Yan and Y. Li

[124] Da Lei, Min Xu, and Shuaian Wang. 2024. A conditional diffusion model for probabilistic estimation of traffic states
at sensor-free locations. Transportation Research Part C: Emerging Technologies 166 (2024), 104798.
[125] Chaoneng Li, Guanwen Feng, Yunan Li, Ruyi Liu, Qiguang Miao, and Liang Chang. 2024. DiffTAD: Denoising diffu-
sion probabilistic models for vehicle trajectory anomaly detection. Knowledge-Based Systems 286 (2024), 111387.
[126] Chenghao Li, Chaoning Zhang, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and
Choong Seon Hong. 2023. Generative AI meets 3D: A survey on text-to-3D in AIGC era. arXiv:2305.06131. Retrieved
from https://arxiv.org/abs/2305.06131
[127] Fuxian Li, Huan Yan, Guangyin Jin, Yue Liu, Yong Li, and Depeng Jin. 2022. Automated spatio-temporal synchronous
modeling with multiple graphs for traffic prediction. In Proceedings of the 31st ACM International Conference on
Information & Knowledge Management. 1084–1093.
[128] Fuxian Li, Huan Yan, Hongjie Sui, Deng Wang, Fan Zuo, Yue Liu, Yong Li, and Depeng Jin. 2023. Periodic shift and
event-aware spatio-temporal graph convolutional network for traffic congestion prediction. In Proceedings of the 31st
ACM International Conference on Advances in Geographic Information Systems. 1–10.
[129] Haitao Li, Qian Cao, Qiaowen Bai, Zhihui Li, and Hongyu Hu. 2023. Multistate time series imputation using genera-
tive adversarial network with applications to traffic data. Neural Computing and Applications 35, 9 (2023), 6545–6567.
[130] Jinyuan Li, Hao Li, Guorong Cui, Yan Kang, Yang Hu, and Yingnan Zhou. 2020. GACNet: A generative adversarial
capsule network for regional epitaxial traffic flow prediction. Computers, Materials & Continua 64, 2 (2020), 925–940.
[131] Jinlong Li, Ruonan Li, Zilin Huang, Pan Wu, and Lunhui Xu. 2023. Dynamic adaptive generative adversarial net-
works with multi-view temporal factorizations for hybrid recovery of missing traffic data. Neural Computing and
Applications 35, 10 (2023), 7677–7696.
[132] Jinning Li, Liting Sun, Jianyu Chen, Masayoshi Tomizuka, and Wei Zhan. 2021. A safe hierarchical planning frame-
work for complex driving scenarios based on reinforcement learning. In 2021 IEEE International Conference on Robot-
ics and Automation (ICRA’21). IEEE, 2660–2666.
[133] Jianxue Li, Yang Xiao, Jiawei Wu, Yaozhi Chen, and Jun Liu. 2022. Attentive dual-head spatial-temporal generative
adversarial networks for crowd flow generation. In 2022 IEEE 33rd Annual International Symposium on Personal,
Indoor and Mobile Radio Communications (PIMRC’22). IEEE, 800–806.
[134] Lincan Li, Jichao Bi, Kaixiang Yang, and Fengji Luo. 2022. Spatial-temporal semantic generative adversarial net-
works for flexible multi-step urban flow prediction. In International Conference on Artificial Neural Networks. Springer,
763–775.
[135] Lincan Li, Jichao Bi, Kaixiang Yang, Fengji Luo, and Luxing Yang. 2022. MGC-GAN: Multi-graph convolutional gen-
erative adversarial networks for accurate citywide traffic flow prediction. In 2022 IEEE International Conference on
Systems, Man, and Cybernetics (SMC’22). IEEE, 2557–2562.
[136] Tao Li, Xu Han, Jiaqi Ma, Marilia Ramos, and Changju Lee. 2023. Operational safety of automated and human driving
in mixed traffic environments: A perspective of car-following behavior. Proceedings of the Institution of Mechanical
Engineers, Part O: Journal of Risk and Reliability 237, 2 (2023), 355–366.
[137] Tianyi Li, Mingfeng Shang, Shian Wang, Matthew Filippelli, and Raphael Stern. 2022. Detecting stealthy cyberattacks
on automated vehicles via generative adversarial networks. In 2022 IEEE 25th International Conference on Intelligent
Transportation Systems (ITSC’22). IEEE, 3632–3637.
[138] Wenli Li, Mengxin Wang, Lingxi Li, Kan Wang, and Yuanzhi Hu. 2024. Game-generative adversarial imitation learn-
ing for pedestrian simulation during pedestrian-vehicle interaction. IEEE Transactions on Intelligent Vehicles (2024),
1–12.
[139] Xiao Li, Guy Rosman, Igor Gilitschenski, Cristian-Ioan Vasile, Jonathan A. DeCastro, Sertac Karaman, and Daniela
Rus. 2021. Vehicle trajectory prediction using generative adversarial network with temporal logic syntax tree features.
IEEE Robotics and Automation Letters 6, 2 (2021), 3459–3466.
[140] Yaochen Li, Xiao Wu, Danhui Lu, Ling Li, Yuehu Liu, and Li Zhu. 2020. Style transfer of urban road images using
generative adversarial networks with structural details. IEEE MultiMedia 27, 3 (2020), 54–65.
[141] Zhonghang Li, Lianghao Xia, Yong Xu, and Chao Huang. 2023. GPT-ST: generative pre-training of spatio-temporal
graph neural networks. Advances in Neural Information Processing Systems 36 (2023), 70229–70246.
[142] Zhimin Li, Haifeng Zheng, and Xinxin Feng. 2018. 3D convolutional generative adversarial networks for missing
traffic data completion. In 2018 10th International Conference on Wireless Communications and Signal Processing
(WCSP’18). IEEE, 1–6.
[143] Yunyi Liang, Zhiyong Cui, Yu Tian, Huimiao Chen, and Yinhai Wang. 2018. A deep generative adversarial archi-
tecture for network-wide spatial-temporal traffic-state estimation. Transportation Research Record 2672, 45 (2018),
87–105.
[144] Zeguang Liao, Han Cheng, Xuan Wang, Xin Tao, Yihuan Zhang, Yifan Dai, and Keqiang Li. 2022. ITGAN: An in-
teractive trajectories generative adversarial network model for automated driving scenario generation. In Society of
Automotive Engineers (SAE)-China Congress. Springer, 554–566.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:37

[145] S. Lima, S. Barbosa, P. Palmeira, L. Matos, I. Secundo, and R. Nascimento. 2017. Systematic review: Techniques and
methods of urban monitoring in intelligent transport systems. ICWMC 2017 17 (2017), 9.
[146] Hongyi Lin, Yang Liu, Shen Li, and Xiaobo Qu. 2023. How generative adversarial networks promote the development
of intelligent transportation systems: A survey. IEEE/CAA Journal of Automatica Sinica 10, 9 (2023), 1781–1796.
[147] Lin Lin, Jiwon Kim, and Sanghyung Ahn. 2021. Car following modelling with constrained generative adversarial
imitation learning. Australasian Transport Research Forum (ATRF). 1–5.
[148] Yilun Lin, Xingyuan Dai, Li Li, and Fei-Yue Wang. 2018. Pattern sensitive prediction of traffic flow based on generative
adversarial framework. IEEE Transactions on Intelligent Transportation Systems 20, 6 (2018), 2395–2400.
[149] Yi Lin, Linchao Li, Hailong Jing, Bin Ran, and Dongye Sun. 2020. Automated traffic incident detection with a smaller
dataset based on generative adversarial networks. Accident Analysis & Prevention 144 (2020), 105628.
[150] Yongjie Lin, Qihang Li, Duanya Lyu, and Xiaofei Wang. 2022. A review of Wi-Fi-based traffic detection technology
in the field of intelligent transportation systems. Buildings 12, 4 (2022), 428.
[151] Yan Lin, Huaiyu Wan, Jilin Hu, Shengnan Guo, Bin Yang, Youfang Lin, and Christian S. Jensen. 2023. Origin-
destination travel time oracle for map-based services. Proceedings of the ACM on Management of Data 1, 3 (2023),
1–27.
[152] Chang Liu, Huan Yan, Hongjie Sui, Haomin Wen, Yuan Yuan, Yuyang Han, Hongsen Liao, Xuetao Ding, Jinghua
Hao, and Yong Li. 2025. MRGRP: Empowering courier route prediction in food delivery service with multi-relational
graph. In The World Wide Web Conference.
[153] Jiaqi Liu, Shiyu Fang, Xuekai Liu, Lulu Guo, Peng Hang, and Jian Sun. 2024. A decision-making GPT model aug-
mented with entropy regularization for autonomous vehicles. 2024 8th CAA International Conference on Vehicular
Control and Intelligence (CVCI). IEEE, 1–6.
[154] Jiaqi Liu, Peng Hang, Xiao Qi, Jianqiang Wang, and Jian Sun. 2023. Mtd-gpt: A multi-task decision-making gpt
model for autonomous driving at unsignalized intersections. In 2023 IEEE 26th International Conference on Intelligent
Transportation Systems (ITSC’23). IEEE, 5154–5161.
[155] Jiaqi Liu, Peng Hang, Xiaocong Zhao, Jianqiang Wang, and Jian Sun. 2024. DDM-lag: A diffusion-based decision-
making model for autonomous vehicles with lagrangian safety enhancement. IEEE Transactions on Artificial Intelli-
gence 6, 3 (2024), 780–791.
[156] Kai Liu and Hongbo Zhang. 2022. Attention based spatio-temporal generative adversarial network for sparse
traffic forecasting. In International Conference on Cloud Computing, Performance Computing, and Deep Learning
(CCPCDL’22), Vol. 12287. SPIE, 492–497.
[157] Mingzhe Liu, Han Huang, Hao Feng, Leilei Sun, Bowen Du, and Yanjie Fu. 2023. Pristi: A conditional diffusion
framework for spatiotemporal imputation. 2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE,
1927–1939.
[158] Ziming Liu, Di Luo, Yilun Xu, Tommi Jaakkola, and Max Tegmark. 2023. GenPhys: From physical processes to gen-
erative models. arXiv:2304.02637. Retrieved from https://arxiv.org/abs/2304.02637
[159] Qingyue Long, Can Rong, Huandong Wang, and Yong Li. 2025. One fits all: General mobility trajectory modeling
via masked conditional diffusion. arXiv:2501.13347. Retrieved from https://arxiv.org/abs/2501.13347
[160] Shuanghu Luo, Ling Yu, Zhongqin Bi, and Yongbin Li. 2020. Traffic sign detection and recognition for intelligent
transportation systems: A survey. Journal of Internet Technology 21, 6 (2020), 1773–1784.
[161] Yisheng Lv. 2023. Artificial intelligence-generated content in intelligent transportation systems: Learning to copy,
change, and create![Editor’s Column]. IEEE Intelligent Transportation Systems Magazine 15, 5 (2023), 2–3.
[162] Lijing Ma, Shiru Qu, Lijun Song, Zhiteng Zhang, and Jie Ren. 2023. A physics-informed generative car-following
model for connected autonomous vehicles. Entropy 25, 7 (2023), 1050.
[163] Shuo Ma, Ouri Wolfson, and Jie Lin. 2011. A survey on trust management for intelligent transportation system. In
Proceedings of the 4th ACM SIGSPATIAL International Workshop on Computational Transportation Science. 18–23.
[164] Athanasios Maimaris and George Papageorgiou. 2016. A review of intelligent transportation systems from a com-
munications technology perspective. In 2016 IEEE 19th International Conference on Intelligent Transportation Systems
(ITSC’16). IEEE, 54–59.
[165] Cheuk Ki Man, Mohammed Quddus, Athanasios Theofilatos, Rongjie Yu, and Marianna Imprialou. 2022. Wasserstein
generative adversarial network to address the imbalanced data problem in real-time crash risk prediction. IEEE
Transactions on Intelligent Transportation Systems 23, 12 (2022), 23002–23013.
[166] Ratna Mandal, Ankita Mandal, Soumi Dutta, Munshi Yusuf Alam, Sujoy Saha, and Subrata Nandi. 2022. Framework
of intelligent transportation system: A survey. In Proceedings of International Conference on Frontiers in Computing
and Systems: COMSYS 2021. Springer, 93–108.
[167] Riccardo Mangiaracina, Alessandro Perego, Giulio Salvadori, and Angela Tumino. 2017. A comprehensive view of
intelligent transport systems for urban smart mobility. International Journal of Logistics Research and Applications
20, 1 (2017), 39–52.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:38 H. Yan and Y. Li

[168] Renata Maria Marè, Claudio Luiz Marte, and Carlos Eduardo Cugnasca. 2016. Visible light communication applied
to intelligent transport systems: An overview. IEEE Latin America Transactions 14, 7 (2016), 3199–3207.
[169] Elezabeth Mathew. 2020. Swarm intelligence for intelligent transport systems: Opportunities and challenges. Swarm
Intelligence for Resource Management in Internet of Things (2020), 131–145.
[170] Matthew McNaughton, Chris Urmson, John M. Dolan, and Jin-Woo Lee. 2011. Motion planning for autonomous
driving with a conformal spatiotemporal lattice. In 2011 IEEE International Conference on Robotics and Automation.
IEEE, 4889–4895.
[171] Silin Meng, Yiwei Wang, Cheng-Fu Yang, Nanyun Peng, and Kai-Wei Chang. 2024. LLM-A*: Large language model
enhanced incremental heuristic search on path planning. arXiv:2407.02511. Retrieved from https://arxiv.org/abs/2407.
02511
[172] Zhaobin Mo, Yongjie Fu, and Xuan Di. 2022. Quantifying uncertainty in traffic state estimation using generative
adversarial networks. In 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC’22). IEEE,
2769–2774.
[173] Rola Naja. 2013. A survey of communications for intelligent transportation systems. In Wireless Vehicular Networks
for Car Collision Avoidance. Springer, 3–35.
[174] Hasan A. H. Naji, Qingji Xue, Huijun Zhu, and Tianfeng Li. 2021. Forecasting taxi demands using generative adver-
sarial networks with multi-source data. Applied Sciences 11, 20 (2021), 9675.
[175] Mahima Nama, Ankita Nath, Nancy Bechra, Jitendra Bhatia, Sudeep Tanwar, Manish Chaturvedi, and Balqies Sadoun.
2021. Machine learning-based traffic scheduling techniques for intelligent transportation system: Opportunities and
challenges. International Journal of Communication Systems 34, 9 (2021), e4814.
[176] Robayet Nasim and Andreas Kassler. 2012. Distributed architectures for intelligent transport systems: A survey. In
2012 Second Symposium on Network Cloud Computing and Applications. IEEE, 130–136.
[177] Marion Neumeier, Michael Botsch, Andreas Tollkühn, and Thomas Berberich. 2021. Variational autoencoder-based
vehicle trajectory prediction with an interpretable latent space. In 2021 IEEE International Intelligent Transportation
Systems Conference (ITSC’21). IEEE, 820–827.
[178] Judith Nkechinyere Njoku, Cosmas Ifeanyi Nwakanma, Gabriel Chukwunonso Amaizu, and Dong-Seong Kim. 2023.
Prospects and challenges of Metaverse application in data-driven intelligent transportation systems. IET Intelligent
Transport Systems 17, 1 (2023), 1–21.
[179] Geunseob Oh and Huei Peng. 2022. Cvae-h: Conditionalizing variational autoencoders via hypernetworks and tra-
jectory forecasting for autonomous driving. arXiv:2201.09874. Retrieved from https://arxiv.org/abs/2201.09874
[180] Anil Ozturk, Mustafa Burak Gunel, Melih Dal, Ugur Yavas, and Nazim Kemal Ure. 2020. Development of a stochastic
traffic environment with generative time-series models for improving generalization capabilities of autonomous
driving agents. In 2020 IEEE Intelligent Vehicles Symposium (IV’20). IEEE, 1343–1348.
[181] Aoyu Pang, Maonan Wang, Man-On Pun, Chung Shue Chen, and Xi Xiong. 2024. iLLM-TSC: Integration reinforce-
ment learning and large language model for traffic signal control policy improvement. arXiv:2407.06025. Retrieved
from https://arxiv.org/abs/2407.06025
[182] Palak Patel, Zunnun Narmawala, and Ankit Thakkar. 2019. A survey on intelligent transportation system using
internet of things. Emerging Research in Computing, Information, Communication and Applications: ERCICA 2018,
Volume 1 (2019), 231–240.
[183] Prashant W. Patil, Akshay Dudhane, and Subrahmanyam Murala. 2020. End-to-end recurrent generative adver-
sarial network for traffic and surveillance applications. IEEE Transactions on Vehicular Technology 69, 12 (2020),
14550–14562.
[184] Mingxing Peng, Kehua Chen, Xusen Guo, Qiming Zhang, Hongliang Lu, Hui Zhong, Di Chen, Meixin Zhu, and
Hai Yang. 2024. Diffusion models for intelligent transportation systems: A survey. arXiv:2409.15816. Retrieved from
https://arxiv.org/abs/2409.15816
[185] Wenchuang Peng, Youfang Lin, Shengnan Guo, Weiwen Tang, Le Liu, and Huaiyu Wan. 2023. Generative-contrastive-
attentive spatial-temporal network for traffic data imputation. In Pacific-Asia Conference on Knowledge Discovery and
Data Mining. Springer, 45–56.
[186] Tsarina Dwi Putri et al. 2021. Intelligent transportation systems (ITS): A systematic review using a Natural Language
Processing (NLP) approach. Heliyon 7, 12 (2021).
[187] Hua Xuan Qin and Pan Hui. 2023. Empowering the metaverse with generative AI: Survey and future directions. 2023
IEEE 43rd International Conference on Distributed Computing Systems Workshops (ICDCSW). IEEE, 85–90.
[188] Yuning Qiu, Teruhisa Misu, and Carlos Busso. 2019. Driving anomaly detection with conditional generative adver-
sarial network using physiological and can-bus data. In 2019 International Conference on Multimodal Interaction.
164–173.
[189] Kashif Naseer Qureshi and Abdul Hanan Abdullah. 2013. A survey on intelligent transportation systems. Middle-East
Journal of Scientific Research 15, 5 (2013), 629–642.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:39

[190] Saeed Rahmani, Asiye Baghbani, Nizar Bouguila, and Zachary Patterson. 2023. Graph neural networks for intelligent
transportation systems: A survey. IEEE Transactions on Intelligent Transportation Systems 24, 8 (2023), 8846–8885.
[191] Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. 2021. Autoregressive denoising diffusion models
for multivariate probabilistic time series forecasting. In International Conference on Machine Learning. PMLR, 8857–
8868.
[192] Shashi Ravi and Mohan Rao Mamdikar. 2022. A review on ITS (Intelligent Transportation Systems) technology. In
2022 International Conference on Applied Artificial Intelligence and Computing (ICAAIC’22). IEEE, 155–159.
[193] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. 2023.
Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition. 13756–13766.
[194] Danilo Rezende and Shakir Mohamed. 2015. Variational inference with normalizing flows. In International Conference
on Machine Learning. PMLR, 1530–1538.
[195] Benjamin B. Rhoades and James M. Conrad. 2017. A survey of alternate methods and implementations of an intelli-
gent transportation system. In SoutheastCon 2017. IEEE, 1–8.
[196] Can Rong, Jingtao Ding, Zhicheng Liu, and Yong Li. 2023. Complexity-aware large scale origin-destination network
generation via diffusion model. arXiv:2306.04873. Retrieved from https://arxiv.org/abs/2306.04873
[197] Can Rong, Huandong Wang, and Yong Li. 2023. Origin-destination network generation via gravity-guided GAN.
arXiv:2306.03390. Retrieved from https://arxiv.org/abs/2306.03390
[198] Debaditya Roy, Tetsuhiro Ishizaka, C Krishna Mohan, and Atsushi Fukuda. 2019. Vehicle trajectory prediction at
intersections using interaction based generative adversarial networks. In 2019 IEEE Intelligent Transportation Systems
Conference (ITSC’19). IEEE, 2318–2323.
[199] Bo-Kai Ruan, Hao-Tang Tsui, Yung-Hui Li, and Hong-Han Shuai. 2024. Traffic scene generation from natural lan-
guage description for autonomous vehicles with large language model. arXiv:2409.09575. Retrieved from https:
//arxiv.org/abs/2409.09575
[200] Sandeep Saharan, Seema Bawa, and Neeraj Kumar. 2020. Dynamic pricing techniques for Intelligent Transportation
System in smart cities: A systematic review. Computer Communications 150 (2020), 603–625.
[201] Kelathodi Kumaran Santhosh, Debi Prosad Dogra, Partha Pratim Roy, and Adway Mitra. 2021. Vehicular trajectory
classification and traffic anomaly detection in videos using a hybrid CNN-VAE Architecture. IEEE Transactions on
Intelligent Transportation Systems 23, 8 (2021), 11891–11902.
[202] Sergio Saponara, Maria Sabrina Greco, and Fulvio Gini. 2019. Radar-on-chip/in-package in autonomous driving ve-
hicles and intelligent transport systems: Opportunities and challenges. IEEE Signal Processing Magazine 36, 5 (2019),
71–84.
[203] Divya Saxena and Jiannong Cao. 2019. D-GAN: Deep generative adversarial nets for spatio-temporal prediction.
arXiv:1907.08556. Retrieved from https://arxiv.org/abs/1907.08556
[204] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei
Zhan, and Mingyu Ding. 2023. Languagempc: Large language models as decision makers for autonomous driving.
arXiv:2310.03026. Retrieved from https://arxiv.org/abs/2310.03026
[205] Khaled Shaaban, Md Hosne Mobarok Shamim, and Khadija Abdur-Rouf. 2021. Visible light communication for intel-
ligent transportation systems: A review of the latest technologies. Journal of Traffic and Transportation Engineering
(English Edition) 8, 4 (2021), 483–492.
[206] Shivani Sharma and Sateesh Kumar Awasthi. 2022. Introduction to intelligent transportation system: Overview,
classification based on physical architecture, and challenges. International Journal of Sensor Networks 38, 4 (2022),
215–240.
[207] Guojiang Shen, Nali Liu, Yinghui Liu, Wenfeng Zhou, and Xiangjie Kong. 2022. Traffic flow imputation based on
multi-perspective spatiotemporal generative adversarial networks. In Proceedings of CECNet 2022. IOS Press, 62–73.
[208] Arun Kumar Singh, Jatan Shrestha, and Nicola Albarella. 2023. Bi-level optimization augmented with conditional
variational autoencoder for autonomous driving in dense traffic. In 2023 IEEE 19th International Conference on Au-
tomation Science and Engineering (CASE’23). IEEE, 1–8.
[209] Bhupendra Singh and Ankit Gupta. 2015. Recent trends in intelligent transportation systems: A review. Journal of
Transport Literature 9 (2015), 30–34.
[210] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. Deep unsupervised learning
using nonequilibrium thermodynamics. In International Conference on Machine Learning. PMLR, 2256–2265.
[211] Xiaozhuang Song, Chenhan Zhang, and J. Q. James. 2021. Learn travel time distribution with graph deep learning
and generative adversarial network. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC’21).
IEEE, 1385–1390.
[212] Zhaoxin Su, Gang Huang, Sanyuan Zhang, and Wei Hua. 2022. Crossmodal transformer based generative framework
for pedestrian trajectory prediction. In 2022 International Conference on Robotics and Automation (ICRA’22). IEEE,
2337–2343.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:40 H. Yan and Y. Li

[213] S. Suhas, V. Vismaya Kalyan, Manoj Katti, B. V. Ajay Prakash, and C. Naveena. 2017. A comprehensive review on
traffic prediction for intelligent transport system. In 2017 International Conference on Recent Advances in Electronics
and Communication Technology (ICRAECT’17). IEEE, 138–143.
[214] Hongjie Sui, Huan Yan, Tianyi Zheng, Wenzhen Huang, Yunlin Zhuang, and Yong Li. 2024. Congestion-aware spatio-
temporal graph convolutional network based A* search algorithm for fastest route search. ACM Transactions on
Knowledge Discovery from Data 18, 7 (2024), 1–19.
[215] Shuo Sun, Zekai Gu, Tianchen Sun, Jiawei Sun, Chengran Yuan, Yuhang Han, Dongen Li, and Marcelo H Ang Jr.
2023. DriveSceneGen: Generating diverse and realistic driving scenarios from scratch. arXiv:2309.14685. Retrieved
from https://arxiv.org/abs/2309.14685
[216] Yasheng Sun, Tao He, Jie Hu, Haiqing Huang, and Biao Chen. 2019. Intent-aware conditional generative adversarial
network for pedestrian path prediction. In 2019 IEEE International Conference on Artificial Intelligence and Computer
Applications (ICAICA’19). IEEE, 155–160.
[217] Yanfeng Sun, Xiangheng Jiang, Yongli Hu, Fuqing Duan, Kan Guo, Boyue Wang, Junbin Gao, and Baocai Yin. 2022.
Dual dynamic spatial-temporal graph convolution network for traffic prediction. IEEE Transactions on Intelligent
Transportation Systems 23, 12 (2022), 23680–23693.
[218] Nikunja K. Swain. 2006. A survey of application of fuzzy logic in intelligent transportation systems (ITS) and rural
ITS. In Proceedings of the IEEE SoutheastCon 2006. IEEE, 85–90.
[219] Yingjuan Tang, Hongwen He, Yong Wang, and Yifan Wu. 2024. Utilizing a diffusion model for pedestrian trajectory
prediction in semi-open autonomous driving environments. IEEE Sensors Journal 24, 10 (2024), 17208–17218.
[220] Jing Tian, Xianmin Song, Pengfei Tao, and Jiahui Liang. 2022. Pattern-adaptive generative adversarial network with
sparse data for traffic state estimation. Physica A: Statistical Mechanics and its Applications 608 (2022), 128254.
[221] Cuong N. N. Tran, Thang Tran Huynh Tat, Vivian W. Y. Tam, and Duc Hoc Tran. 2023. Factors affecting intelligent
transport systems towards a smart city: A critical review. International Journal of Construction Management 23, 12
(2023), 1982–1998.
[222] Hardik Trivedi, Sudeep Tanwar, and Priyank Thakkar. 2019. Software defined network-based vehicular adhoc net-
works for intelligent transportation system: Recent advances and future challenges. In Futuristic Trends in Network
and Communication Technologies: First International Conference, FTNCT 2018, Solan, India, February 9–10, 2018, Revised
Selected Papers 1. Springer, 325–337.
[223] Jingzhi Tu, Gang Mei, and Francesco Piccialli. 2021. Incomplete vehicle information completion using generative
adversarial network to enhance the safety of autonomous driving. In Second International Conference on Industrial
IoT, Big Data, and Supply Chain, Vol. 12128. SPIE, 61–66.
[224] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia
Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems 30 (2017), 1–11.
[225] Matthew Veres and Medhat Moussa. 2019. Deep learning for intelligent transportation systems: A survey of emerging
trends. IEEE Transactions on Intelligent Transportation Systems 21, 8 (2019), 3152–3168.
[226] Chalavadi Vishnu, Vineel Abhinav, Debaditya Roy, C. Krishna Mohan, and Ch Sobhan Babu. 2023. Improving multi-
agent trajectory prediction using traffic states on interactive driving scenarios. IEEE Robotics and Automation Letters
8, 5 (2023), 2708–2715.
[227] Huandong Wang, Yong Li, Depeng Jin, and Zhu Han. 2021. Attentional Markov model for human mobility prediction.
IEEE Journal on Selected Areas in Communications 39, 7 (2021), 2213–2225.
[228] Huandong Wang, Huan Yan, Can Rong, Yuan Yuan, Fenyu Jiang, Zhenyu Han, Hongjie Sui, Depeng Jin, and Yong Li.
2024. Multi-scale simulation of complex systems: A perspective of integrating knowledge and data. Comput. Surveys
56, 12 (2024), 1–38.
[229] Huandong Wang, Qiaohong Yu, Yu Liu, Depeng Jin, and Yong Li. 2021. Spatio-temporal urban knowledge graph
enabled mobility prediction. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 5,
4 (2021), 1–24.
[230] Huandong Wang, Qizhong Zhang, Yuchen Wu, Depeng Jin, Xing Wang, Lin Zhu, and Li Yu. 2023. Synthesizing
human trajectories based on variational point processes. IEEE Transactions on Knowledge and Data Engineering 36, 4
(2023), 1785–1799.
[231] Jiguang Wang, Yilun Zhang, Xinjie Xing, Yuanzhu Zhan, Wai Kin Victor Chan, and Sunil Tiwari. 2022. A data-driven
system for cooperative-bus route planning based on generative adversarial network and metric learning. Annals of
Operations Research 339, 1 (2024), 427–453.
[232] Min Wang, Jianqun Cui, Yew Wee Wong, Yanan Chang, Libing Wu, and Jiong Jin. 2024. Urban vehicle trajectory
generation based on generative adversarial imitation learning. IEEE Transactions on Vehicular Technology 73, 12
(2024), 18237–18249.
[233] Meng Wang, Serge P. Hoogendoorn, Winnie Daamen, Bart van Arem, and Riender Happee. 2015. Game theoretic ap-
proach for predictive lane-changing and car-following control. Transportation Research Part C: Emerging Technologies
58 (2015), 73–92.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:41

[234] Maonan Wang, Aoyu Pang, Yuheng Kan, Man-On Pun, Chung Shue Chen, and Bo Huang. 2024. LLM-assisted light:
Leveraging large language model capabilities for human-mimetic traffic signal control in complex urban environ-
ments. arXiv:2403.08337. Retrieved from https://arxiv.org/abs/2403.08337
[235] Mudan Wang, Huan Yan, Huandong Wang, Yong Li, and Depeng Jin. 2023. Contagion process guided cross-scale
spatio-temporal graph neural network for traffic congestion prediction. In Proceedings of the 31st ACM International
Conference on Advances in Geographic Information Systems. 1–11.
[236] Qi Wang. 2022. VARL: A variational autoencoder-based reinforcement learning framework for vehicle routing prob-
lems. Applied Intelligence 52 (2022), 1–14.
[237] Tong Wang, Xiaodan Wang, Ziping Cui, Yue Cao, and Chakkaphong Suthaputchakun. 2019. Survey on cooperatively
V2X downloading for intelligent transport systems. IET Intelligent Transport Systems 13, 1 (2019), 13–21.
[238] Tao Wang, Yushu Zhang, Shuren Qi, Ruoyu Zhao, Zhihua Xia, and Jian Weng. 2023. Security and privacy on gener-
ative data in aigc: A survey. arXiv:2309.09435. Retrieved from https://arxiv.org/abs/2309.09435
[239] Xu Wang, Xiaoming Chen, and Yanping Wang. 2021. Small vehicle classification in the wild using generative adver-
sarial network. Neural Computing and Applications 33 (2021), 5369–5379.
[240] Yuntao Wang, Yanghe Pan, Miao Yan, Zhou Su, and Tom H. Luan. 2023. A survey on ChatGPT: AI-Generated contents,
challenges, and solutions. arXiv:2305.18339. Retrieved from https://arxiv.org/abs/2305.18339
[241] Yun-Cheng Wang, Jintang Xue, Chengwei Wei, and C.-C. Jay Kuo. 2023. An overview on generative AI at scale with
edge-cloud computing. IEEE Open Journal of the Communications Society 4 (2023), 2952–2971.
[242] Zixin Wang, Hanyu Zhu, Mingcheng He, Yong Zhou, Xiliang Luo, and Ning Zhang. 2021. GAN and multi-agent DRL
based decentralized traffic light signal control. IEEE Transactions on Vehicular Technology 71, 2 (2021), 1333–1348.
[243] Chuheng Wei, Guoyuan Wu, Matthew J. Barth, Amr Abdelraouf, Rohit Gupta, and Kyungtae Han. 2024. KI-GAN:
Knowledge-informed generative adversarial networks for enhanced multi-vehicle trajectory forecasting at signalized
intersections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7115–7124.
[244] Junqing Wei, Jarrod M. Snider, Tianyu Gu, John M. Dolan, and Bakhtiar Litkouhi. 2014. A behavioral planning frame-
work for autonomous driving. In 2014 IEEE Intelligent Vehicles Symposium Proceedings. IEEE, 458–464.
[245] Haomin Wen, Youfang Lin, Yutong Xia, Huaiyu Wan, Roger Zimmermann, and Yuxuan Liang. 2023. Diffstg: Proba-
bilistic spatio-temporal graph forecasting with denoising diffusion models. Proceedings of the 31st ACM International
Conference on Advances in Geographic Information Systems. 1–12.
[246] Nurhadi Wijaya, Sri Hasta Mulyani, and Albertus Christian Noviadi Prabowo. 2022. DeepDrive: Effective distracted
driver detection using generative adversarial networks (GAN) algorithm. Iran Journal of Computer Science 5, 3 (2022),
221–227.
[247] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Hong Lin. 2023. Ai-generated content (aigc): A survey.
arXiv:2304.06632. Retrieved from https://arxiv.org/abs/2304.06632
[248] Lan Wu, Tian Gao, Chenglin Wen, Kunpeng Zhang, and Fanshi Kong. 2021. A high-dimensional video sequence com-
pletion method with traffic data completion generative adversarial networks. Wireless Communications and Mobile
Computing 2021, 1 (2021), 1–9.
[249] Lan Wu, Han Wang, Tian Gao, Binquan Li, and Fanshi Kong. 2022. A traffic video completion model based on
generative adversarial networks. In Proceedings of 2021 Chinese Intelligent Automation Conference. Springer, 658–669.
[250] Zhu Xiao, Jinmei Shu, Hongbo Jiang, Geyong Min, Hongyang Chen, and Zhu Han. 2022. Perception task offloading
with collaborative computation for autonomous driving. IEEE Journal on Selected Areas in Communications 41, 2
(2022), 457–473.
[251] Lei Xie, Tao Guo, Jiliang Chang, Chengpeng Wan, Xinyuan Hu, Yang Yang, and Changkui Ou. 2023. A novel model for
ship trajectory anomaly detection based on gaussian mixture variational autoencoder. IEEE Transactions on Vehicular
Technology 72, 11 (2023), 13826–13835.
[252] Hao Xing, Jianming Hu, and Zuo Zhang. 2022. Multi-modal vehicle trajectory prediction via attention-based condi-
tional variational autoencoder. In 2022 37th Youth Academic Annual Conference of Chinese Association of Automation
(YAC’22). IEEE, 1367–1373.
[253] Gang Xiong, Zhishuai Li, Meihua Zhao, Yu Zhang, Qinghai Miao, Yisheng Lv, and Fei-Yue Wang. 2023. TrajSGAN:
A semantic-guiding adversarial network for urban trajectory generation. IEEE Transactions on Computational Social
Systems 11, 2 (2023), 1733–1743.
[254] Zhang Xiong, Hao Sheng, WenGe Rong, and Dave E. Cooper. 2012. Intelligent transportation systems for smart cities:
A progress review. Science China Information Sciences 55 (2012), 2908–2914.
[255] Chejian Xu, Ding Zhao, Alberto Sangiovanni-Vincentelli, and Bo Li. 2023. DiffScene: Diffusion-based safety-critical
scenario generation for autonomous vehicles. In The Second Workshop on New Frontiers in Adversarial Machine
Learning.
[256] Dongwei Xu, Peng Peng, Chenchen Wei, Defeng He, and Qi Xuan. 2020. Road traffic network state prediction based
on a generative adversarial network. IET Intelligent Transport Systems 14, 10 (2020), 1286–1294.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:42 H. Yan and Y. Li

[257] Dongwei Xu, Chenchen Wei, Peng Peng, Qi Xuan, and Haifeng Guo. 2020. GE-GAN: A novel deep learning frame-
work for road traffic state estimation. Transportation Research Part C: Emerging Technologies 117 (2020), 102635.
[258] Dongwei Xu, Zefeng Yu, Xiangwang Liao, and Haifeng Guo. 2024. A graph deep reinforcement learning traffic signal
control for multiple intersections considering missing data. IEEE Transactions on Vehicular Technology 73, 12 (2024),
18307–18319.
[259] Dongwei Xu, Zefeng Yu, Tian Tian, and Yanfang Yang. 2022. Generative adversarial network for imputation of road
network traffic state data. In China National Conference on Big Data and Social Computing. Springer, 80–96.
[260] Minrui Xu, Hongyang Du, Dusit Niyato, Jiawen Kang, Zehui Xiong, Shiwen Mao, Zhu Han, Abbas Jamalipour,
Dong In Kim, Victor Leung, et al. 2023. Unleashing the power of edge-cloud generative ai in mobile networks: A
survey of aigc services. arXiv:2303.16129. Retrieved from https://arxiv.org/abs/2303.16129
[261] Minrui Xu, Dusit Niyato, Junlong Chen, Hongliang Zhang, Jiawen Kang, Zehui Xiong, Shiwen Mao, and Zhu Han.
2023. Generative AI-empowered simulation for autonomous driving in vehicular mixed reality metaverses. IEEE
Journal of Selected Topics in Signal Processing 17, 5 (2023), 1064–1079.
[262] Xovee Xu, Yutao Wei, Pengyu Wang, Xucheng Luo, Fan Zhou, and Goce Trajcevski. 2023. Diffusion probabilistic
modeling for fine-grained urban traffic flow inference with relaxed structural constraint. In ICASSP 2023-2023 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP’23). IEEE, 1–5.
[263] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao.
2024. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automa-
tion Letters 9, 10 (2024), 8186–8193.
[264] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. 2023. Feature prediction diffusion model for
video anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 5527–5537.
[265] Huan Yan, Guangyin Jin, Deng Wang, Yue Liu, and Yong Li. 2022. Jointly modeling intersections and road segments
for travel time estimation via dual graph convolutional networks. In International Conference on Spatial Data and
Intelligence. Springer, 19–34.
[266] Tijin Yan, Hongwei Zhang, Tong Zhou, Yufeng Zhan, and Yuanqing Xia. 2021. Scoregrad: Multivariate probabilistic
time series forecasting with continuous energy-based generative models. arXiv:2106.10121. Retrieved from https:
//arxiv.org/abs/2106.10121
[267] Bing Yang, Yan Kang, YaoYao Yuan, Xin Huang, and Hao Li. 2021. ST-LBAGAN: Spatio-temporal learnable bidirec-
tional attention generative adversarial networks for missing traffic data imputation. Knowledge-Based Systems 215
(2021), 106705.
[268] Bing Yang, Yan Kang, Yaoyao Yuan, Hao Li, and Fei Wang. 2022. ST-FVGAN: Filling series traffic missing values with
generative adversarial network. Transportation Letters 14, 4 (2022), 407–415.
[269] Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki.
2024. Diffusion-es: Gradient-free planning with diffusion for autonomous driving and zero-shot instruction following.
arXiv:2402.06559. Retrieved from https://arxiv.org/abs/2402.06559
[270] Chen Yang, Aaron Xuxiang Tian, Dong Chen, Tianyu Shi, and Arsalan Heydarian. 2024. WcDT: World-centric diffu-
sion transformer for traffic scene generation. arXiv:2404.02082. Retrieved from https://arxiv.org/abs/2404.02082
[271] Zan Yang, Wei Nai, Dan Li, Lu Liu, and Ziyu Chen. 2024. A mixed generative adversarial imitation learning based
vehicle path planning algorithm. IEEE Access 12 (2024), 85859–85879.
[272] Zi Yang and Lilian S. C. Pun-Cheng. 2018. Vehicle detection in intelligent transportation systems and its applications
under varying environments: A review. Image and Vision Computing 69 (2018), 143–154.
[273] Huaiyuan Yao, Longchao Da, Vishnu Nandam, Justin Turnau, Zhiwei Liu, Linsey Pang, and Hua Wei. 2025. Comal:
Collaborative multi-agent large language models for mixed-autonomy traffic. Proceedings of the 2025 SIAM Interna-
tional Conference on Data Mining (SDM). Society for Industrial and Applied Mathematics. 409–418.
[274] Peijun Ye, Fenghua Zhu, Yisheng Lv, Xiao Wang, and Yuanyuan Chen. 2022. Efficient calibration of agent-based traffic
simulation using variational auto-encoder. In 2022 IEEE 25th International Conference on Intelligent Transportation
Systems (ITSC’22). IEEE, 3077–3082.
[275] Jinhui Yi, Huan Yan, Haotian Wang, Jian Yuan, and Yong Li. 2024. RCCNet: A spatial-temporal neural network model
for logistics delivery timely rate prediction. ACM Transactions on Intelligent Systems and Technology 15, 6 (2024), 1–21.
[276] Byeonghyeop Yu, Yongjin Lee, and Keemin Sohn. 2020. Forecasting road traffic speeds by considering area-wide
spatio-temporal dependencies based on a graph convolutional neural network (GCN). Transportation Research Part
C: Emerging Technologies 114 (2020), 189–204.
[277] Fudan Yu, Huan Yan, Rui Chen, Guozhen Zhang, Yu Liu, Meng Chen, and Yong Li. 2023. City-scale vehicle trajectory
data from traffic camera videos. Scientific Data 10, 1 (2023), 711.
[278] James Jian Qiao Yu and Jiatao Gu. 2019. Real-time traffic speed estimation with graph convolutional generative
autoencoder. IEEE Transactions on Intelligent Transportation Systems 20, 10 (2019), 3940–3951.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:43

[279] Haitao Yuan and Guoliang Li. 2021. A survey of traffic prediction: From spatio-temporal data to intelligent trans-
portation. Data Science and Engineering 6 (2021), 63–85.
[280] Xinyu Yuan, Yan Qiao, Pei Zhao, Rongyao Hu, and Benchu Zhang. 2023. Traffic matrix estimation based on denoising
diffusion probabilistic model. In 2023 IEEE Symposium on Computers and Communications (ISCC’23). IEEE, 316–322.
[281] Ye Yuan, Yong Zhang, Boyue Wang, Yuan Peng, Yongli Hu, and Baocai Yin. 2022. STGAN: Spatio-temporal generative
adversarial network for traffic data imputation. IEEE Transactions on Big Data 9, 1 (2022), 200–211.
[282] Taeyoung Yun, Haewon Jung, and Jiwoo Son. 2023. Imputation as inpainting: Diffusion models for SpatioTemporal
data imputation. https://openreview.net/pdf?id=QUANtQnx30l [Access date: 05/30/2025]
[283] Y. Yun, D. Jeong, and S. Lim. 2019. Data-driven human-like cut-in driving model using generative adversarial network.
Electronics Letters 55, 24 (2019), 1288–1290.
[284] Mohsen Zand, Ali Etemad, and Michael Greenspan. 2023. Flow-based spatio-temporal structured prediction of dy-
namics. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 11 (2023), 13523–13535.
[285] Di Zang, Yang Fang, Zhihua Wei, Keshuang Tang, and Jiujun Cheng. 2019. Traffic flow data prediction using residual
deconvolution based deep generative network. IEEE Access 7 (2019), 71311–71322.
[286] Mohammad Zarei and Bruce Hellinga. 2021. Crash data augmentation using conditional generative adversarial net-
works (CGAN) for improving safety performance functions. arXiv:2112.12263. Retrieved from https://arxiv.org/abs/
2112.12263
[287] Aditi Zear, Pradeep Kumar Singh, and Yashwant Singh. 2016. Intelligent transport system: A progressive review.
(2016).
[288] Chaoning Zhang, Chenshuang Zhang, Chenghao Li, Yu Qiao, Sheng Zheng, Sumit Kumar Dam, Mengchun Zhang,
Jung Uk Kim, Seong Tae Kim, Jinwoo Choi, Gyeong-Moon Park, Sung-Ho Bae, Lik-Hang Lee, Pan Hui, In So Kweon,
and Choong Seon Hong. 2023. One small step for generative ai, one giant leap for agi: A complete survey on chatgpt
in aigc era. arXiv:2304.06488. Retrieved from https://arxiv.org/abs/2304.06488
[289] Chaoning Zhang, Chenshuang Zhang, Sheng Zheng, Yu Qiao, Chenghao Li, Mengchun Zhang, Sumit Kumar Dam,
Chu Myaet Thwal, Ye Lin Tun, Le Luang Huy, Donguk kim, Sung-Ho Bae, Lik-Hang Lee, Yang Yang, Heng Tao Shen,
In So Kweon, and Choong Seon Hong. 2023. A complete survey on generative ai (aigc): Is chatgpt from gpt-4 to gpt-5
all you need? arXiv:2303.11717. Retrieved from https://arxiv.org/abs/2303.11717
[290] He Zhang, Jian Sun, and Ye Tian. 2022. Accelerated testing for highly automated vehicles: A combined method based
on importance sampling and normalizing flows. In 2022 IEEE 25th International Conference on Intelligent Transporta-
tion Systems (ITSC’22). IEEE, 574–579.
[291] Jinlei Zhang, Hua Li, Lixing Yang, Guangyin Jin, Jianguo Qi, and Ziyou Gao. 2023. A spatiotemporal graph generative
adversarial networks for short-term passenger flow prediction in urban rail transit systems. International Journal of
General Systems 52, 6 (2023), 694–721.
[292] Jinlei Zhang, Hua Li, Shuxin Zhang, Lixing Yang, Guangyin Jin, and Jianguo Qi. 2023. A spatiotemporal graph gen-
erative adversarial networks for short-term passenger flow prediction in urban rail transit systems. International
Journal of General Systems 52, 6 (2023), 694–721.
[293] Kunpeng Zhang, Zhengbing He, Liang Zheng, Liang Zhao, and Lan Wu. 2021. A generative adversarial network
for travel times imputation using trajectory data. Computer-Aided Civil and Infrastructure Engineering 36, 2 (2021),
197–212.
[294] Liang Zhang, Jianqing Wu, Jun Shen, Ming Chen, Rui Wang, Xinliang Zhou, Cankun Xu, Quankai Yao, and Qiang Wu.
2021. SATP-GAN: Self-attention based generative adversarial network for traffic flow prediction. Transportmetrica
B: Transport Dynamics 9, 1 (2021), 552–568.
[295] Liming Zhang, Liang Zhao, and Dieter Pfoser. 2022. Factorized deep generative models for end-to-end trajectory
generation with spatiotemporal validity constraints. In Proceedings of the 30th International Conference on Advances
in Geographic Information Systems. 1–12.
[296] Mengchun Zhang, Maryam Qamar, Taegoo Kang, Yuna Jung, Chenshuang Zhang, Sung-Ho Bae, and Chaoning
Zhang. 2023. A survey on graph diffusion models: Generative ai in science for molecule, protein and material.
arXiv:2304.01565. Retrieved from https://arxiv.org/abs/2304.01565
[297] Ruichen Zhang, Ke Xiong, Hongyang Du, Dusit Niyato, Jiawen Kang, Xuemin Shen, and H. Vincent Poor. 2024.
Generative AI-enabled vehicular networks: Fundamentals, framework, and case study. IEEE Network 38, 4 (2024),
259–267.
[298] Siyao Zhang, Daocheng Fu, Wenzhe Liang, Zhao Zhang, Bin Yu, Pinlong Cai, and Baozhen Yao. 2024. Trafficgpt:
Viewing, processing and interacting with traffic foundation models. Transport Policy 150 (2024), 95–105.
[299] Shunyang Zhang, Senzhang Wang, Xianzhen Tan, Ruochen Liu, Jian Zhang, and Jianxin Wang. 2023. sasdim: Self-
adaptive noise scaling diffusion model for spatial time series imputation. arXiv:2309.01988. Retrieved from https:
//arxiv.org/abs/2309.01988
[300] Tong Zhang, Jianlong Wang, and Jie Liu. 2021. A gated generative adversarial imputation approach for signalized
road networks. IEEE Transactions on Intelligent Transportation Systems 23, 8 (2021), 12144–12160.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
315:44 H. Yan and Y. Li

[301] Weibin Zhang, Pulin Zhang, Yinghao Yu, Xiying Li, Salvatore Antonio Biancardo, and Junyi Zhang. 2021. Missing
data repairs for traffic flow with self-attention generative adversarial imputation net. IEEE Transactions on Intelligent
Transportation Systems 23, 7 (2021), 7919–7930.
[302] Xiaoxi Zhang, Yuan Gao, Xin Wang, Jun Feng, and Yan Shi. 2022. GeoSDVA: A semi-supervised dirichlet varia-
tional autoencoder model for transportation mode identification. ISPRS International Journal of Geo-Information 11,
5 (2022), 290.
[303] Xin Zhang, Yanhua Li, Xun Zhou, and Jun Luo. 2020. cgail: Conditional generative adversarial imitation learning–an
application in taxi drivers’ strategy learning. IEEE Transactions on Big Data 8, 5 (2020), 1288–1300.
[304] Xin Zhang, Yanhua Li, Xun Zhou, Ziming Zhang, and Jun Luo. 2020. Trajgail: Trajectory generative adversarial
imitation learning for long-term decision analysis. In 2020 IEEE International Conference on Data Mining (ICDM’20).
IEEE, 801–810.
[305] Yingxue Zhang, Yanhua Li, Xun Zhou, Xiangnan Kong, and Jun Luo. 2019. TrafficGAN: Off-deployment traffic esti-
mation with traffic generative adversarial networks. In 2019 IEEE International Conference on Data Mining (ICDM’19).
IEEE, 1474–1479.
[306] Yingxue Zhang, Yanhua Li, Xun Zhou, Xiangnan Kong, and Jun Luo. 2020. Curb-gan: Conditional urban traffic estima-
tion through spatio-temporal generative adversarial networks. In Proceedings of the 26th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining. 842–852.
[307] Yingxue Zhang, Yanhua Li, Xun Zhou, Xiangnan Kong, and Jun Luo. 2020. Off-deployment traffic estimation–A
traffic generative adversarial networks approach. IEEE Transactions on Big Data 8, 4 (2020), 1084–1095.
[308] Yingxue Zhang, Yanhua Li, Xun Zhou, Xiangnan Kong, and Jun Luo. 2022. STrans-GAN: Spatially-transferable
generative adversarial networks for urban traffic estimation. In 2022 IEEE International Conference on Data Mining
(ICDM’22). IEEE, 743–752.
[309] Yingxue Zhang, Yanhua Li, Xun Zhou, Zhenming Liu, and Jun Luo. 2021. C 3-GAN: Complex-condition-controlled
urban traffic estimation through generative adversarial networks. In 2021 IEEE International Conference on Data
Mining (ICDM’21). IEEE, 1505–1510.
[310] Yuxuan Zhang, Senzhang Wang, Bing Chen, and Jiannong Cao. 2019. GCGAN: Generative adversarial nets with
graph CNN for network-scale traffic prediction. In 2019 International Joint Conference on Neural Networks (IJCNN’19).
IEEE, 1–8.
[311] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. 2025.
Drivedreamer-2: Llm-enhanced world models for diverse driving video generation. Proceedings of the AAAI Con-
ference on Artificial Intelligence 39, 10 (2025), 10412–10420.
[312] Han Zhao, Ruikang Luo, Bowen Yao, Yiyi Wang, Shaoqing Hu, and Rong Su. 2022. GraphSAGE-based generative
adversarial network for short-term traffic speed prediction problem. In 2022 17th International Conference on Control,
Automation, Robotics and Vision (ICARCV’22). IEEE, 837–842.
[313] Jingyuan Zhao, Wenyi Zhao, Bo Deng, Zhenghong Wang, Feng Zhang, Wenxiang Zheng, Wanke Cao, Jinrui Nan,
Yubo Lian, and Andrew F. Burke. 2023. Autonomous driving system: A comprehensive survey. Expert Systems with
Applications 242 (2023), 122836.
[314] Xiaocong Zhao, Ye Tian, and Jian Sun. 2021. Yield or rush? Social-preference-aware driving interaction modeling
using game-theoretic framework. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC’21).
IEEE, 453–459.
[315] Yilin Zhao, Yijie Xun, Jiajia Liu, and Siyu Ma. 2022. GVIDS: A reliable vehicle intrusion detection system based on
generative adversarial network. In GLOBECOM 2022-2022 IEEE Global Communications Conference. IEEE, 4310–4315.
[316] Zihao Zhao, Min Jiang, Jia Guo, Xiaoyu Yang, Yudie Hu, and Xianlong Zhou. 2022. Raindrop removal for in-vehicle
camera images with generative adversarial network. In 2022 IEEE International Conference on Systems, Man, and
Cybernetics (SMC’22). IEEE, 3131–3136.
[317] Hongling Zheng, Xiang Li, Yongfeng Li, Ziqin Yan, and Tinghong Li. 2022. GCN-GAN: Integrating graph convolu-
tional network and generative adversarial network for traffic flow prediction. IEEE Access 10 (2022), 94051–94062.
[318] Ting Zhong, Haoyang Yu, Rongfan Li, Xovee Xu, Xucheng Luo, and Fan Zhou. 2022. Probabilistic fine-grained urban
flow inference with normalizing flows. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP’22). IEEE, 3663–3667.
[319] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray.
2023. Language-guided traffic simulation via scene-level diffusion. Conference on Robot Learning. PMLR, 144–177.
[320] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone.
2023. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics
and Automation (ICRA’23). IEEE, 3560–3566.
[321] Danyang Zhou, Huxiao Wang, Wei Li, Yi Zhou, Nan Cheng, and Ning Lu. 2021. SA-SGAN: A vehicle trajectory pre-
diction model based on generative adversarial networks. In 2021 IEEE 94th Vehicular Technology Conference (VTC2021-
Fall’21). IEEE, 1–5.

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.
Generative AI for Intelligent Transportation Systems: Road Transportation Perspective 315:45

[322] Xuan Zhou, Ruimin Ke, Hao Yang, and Chenxi Liu. 2021. When intelligent transportation systems sensing meets
edge computing: Vision and challenges. Applied Sciences 11, 20 (2021), 9680.
[323] Zhilun Zhou, Jingtao Ding, Yu Liu, Depeng Jin, and Yong Li. 2023. Towards generative modeling of urban flow
through knowledge-enhanced denoising diffusion. Proceedings of the 31st ACM International Conference on Advances
in Geographic Information Systems. 1–12.
[324] Ziai Zhou, Bin Zhou, and Hao Liu. 2024. DynamicRouteGPT: A real-time multi-vehicle dynamic navigation frame-
work based on large language models. arXiv:2408.14185. Retrieved from https://arxiv.org/abs/2408.14185
[325] Li Zhu, Fei Richard Yu, Yige Wang, Bin Ning, and Tao Tang. 2018. Big data analytics in intelligent transportation
systems: A survey. IEEE Transactions on Intelligent Transportation Systems 20, 1 (2018), 383–398.
[326] Yuanshao Zhu, Yongchao Ye, Xiangyu Zhao, and James JQ Yu. 2023. Diffusion model for GPS trajectory generation.
arXiv:2304.11582. Retrieved from https://arxiv.org/abs/2304.11582
[327] Zefang Zong, Huan Yan, Hongjie Sui, Haoxiang Li, Peiqi Jiang, and Yong Li. 2023. An AI-based simulation and
optimization framework for logistic systems. In Proceedings of the 32nd ACM International Conference on Information
and Knowledge Management. 5138–5142.
[328] Jiayu Zou, Kun Tian, Zheng Zhu, Yun Ye, and Xingang Wang. 2024. Diffbev: Conditional diffusion model for bird’s
eye view perception. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 7846–7854.

Received 5 December 2023; revised 9 November 2024; accepted 3 February 2025

ACM Comput. Surv., Vol. 57, No. 12, Article 315. Publication date: July 2025.