This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Date of publication xxxx 00, 0000, date of current version June 19, 2019.
Digital Object Identifier

An Automatic Car Accident Detection

Method Based on Cooperative Vehicle
Infrastructure Systems
DAXIN TIAN1 , (Senior Member, IEEE), CHUANG ZHANG1 , XUTING DUAN1 , AND XIXIAN
WANG1
1
Beijing Advanced Innovation Center for Big Data and Brain Computing, Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems and Safety
Control, School of Transportation Science and Engineering, Beihang University, Beijing 100191, China.
Corresponding author: Xuting Duan (e-mail: duanxuting@buaa.edu.cn).
This research was supported in part by the National Natural Science Foundation of China under Grant Nos. 61672082 and 61822101,
Beijing Municipal Natural Science Foundation Nos. 4181002.

ABSTRACT Car accidents cause a large number of deaths and disabilities every day, a certain proportion
of which result from untimely treatment and secondary accidents. To some extent, automatic car accident
detection can shorten response time of rescue agencies and vehicles around accidents to improve rescue
efficiency and traffic safety level. In this paper, we proposed an automatic car accident detection method
based on Cooperative Vehicle Infrastructure Systems (CVIS) and machine vision. First of all, a novel image
dataset CAD-CVIS is established to improve accuracy of accident detection based on intelligent roadside
devices in CVIS. Especially, CAD-CVIS is consisted of various kinds of accident types, weather conditions
and accident location, which can improve self-adaptability of accident detection methods among different
traffic situations. Secondly, we develop a deep neural network model YOLO-CA based on CAD-CVIS and
deep learning algorithms to detect accident. In the model, we utilize Multi-Scale Feature Fusion (MSFF)
and loss function with dynamic weights to enhance performance of detecting small objects. Finally, our
experiment study evaluates performance of YOLO-CA for detecting car accidents, and the results show
that our proposed method can detect car accident in 0.0461 seconds (21.6FPS) with 90.02% average
precision (AP). In additionally, we compare YOLO-CA with other object detection models, and the results
demonstrate the comprehensive performance improvement on the accuracy and real-time over other models.

INDEX TERMS Car accident detection, CVIS, machine vision, deep learning

I. INTRODUCTION captured by vehicular GPS devices to detect car accident,

CCORDING to the World Health Organization, there such as acceleration and velocity. However, these methods
A are about 1.35 million deaths and 20-50 million in-
juries as a result of the car accident globally every year
based on single type of features cannot meet the performance
need of accident detection in the aspect of accuracy and real-
[1]. Especially, a certain proportion of deaths and injuries time. With the development of computer and communication
are due to untimely treatment and secondary accidents [2], technologies, Cooperative Vehicle Infrastructure System and
which results from that rescue agency and vehicles around Internet of Vehicles have been developed rapidly in recent
accident cannot obtain quick response about the accident [3], years [11]–[13]. Moreover, the image recognition based on
[4]. Therefore, it is vital important to develop an efficient video captured by intelligent roadside devices in CVIS has
accident detection method, which can significantly reduce become one of research hotspots in the field of intelligent
both the number of deaths and injuries as well as the impact transportation system [14], [15]. For traffic situation aware-
and severity of accidents [5]. Under this background, many ness, image recognition technology has advantages of high
fundamental projects and studies to develop efficient detec- efficiency, flexible installation and low maintenance costs.
tion method have been launched for developing and testing Therefore, the image recognition has been applied to detec-
[6]–[10]. tion pedestrian, vehicle, traffic sign and so on successfully
The traditional methods utilize vehicle motion parameters [16]–[20]. In generally, there are many distinctive image and

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

video features in traffic accidents, such as vehicle collision, High Sensitive GPS receiver is considered as the index for
rollover and so on. To some extent, these features can be detecting accidents, and the GSM/GPRS modem is utilized to
used to detect or predict car accidents. Accordingly, some send the location of the accident. [24] presented a prototype
researchers apply the machine vision technology based on system called e-NOTIFY, which monitors the change of ac-
deep-learning into methods of car accident detection. These celeration to detect accident and utilize V2X communication
methods extract and process complex image features instead technologies to report it. To a certain extent, these methods
of single vehicle motion parameter, which improves the can detect and report car accidents in short time, and improve
accuracy of detecting car accidents. However, the datasets the efficiency of car accidents warning. However, the vehi-
of these methods are mostly captured by car cameras or cle running condition before car accidents is complex and
cell phones of pedestrian, which is not suitable for roadside unpredictable, and the accuracy of accident detection only
devices in CVIS. In additionally, the reliability and real-time based on speed and acceleration may be low. In addition, they
performance of these methods need to be improved to meet rely too heavily on vehicular monitoring and communication
the requirements of car accident detection. equipment, which may be unreliable or damaged in some
In this paper, we propose a data-driven car accident detec- extreme circumstances, such as heavy canopy, underground
tion method based on CVIS, whose goal is improving effi- tunnel, and serious car accidents.
ciency and accuracy of car accident response. With the goal,
we focus on such a general application scenario when there B. METHOD BASED VIDEO FEATURES
is an accident on the road, roadside intelligent devices recog- With the development of machine vision and artificial neural
nize and locate it efficiently. First, we build a novel dataset, network technology, more and more applications based on
Car Accident Detection for Cooperative Vehicle Infrastruc- video processing have been applied in transportation and
ture System dataset (CAD-CVIS), which is more suitable for vehicle fields. Under this background, some researchers uti-
car accident detection based on roadside intelligent devices lized video features of the car accident to detect it. [25] pre-
in CVIS. Then, a deep learning model YOLO-CA based on sented a Dynamic-Spatial-Attention Recurrent Neural Net-
CAD-CVIS is developed to detect car accident. Especially, work (RNN) for anticipating accidents in dashcam videos,
we optimize the network of traditional deep learning models which can predict accidents about 2 seconds before they
YOLO [21] to build network of YOLO-CA, which is more occur with 80% recall and 56.14% precision. [26] proposed
accurate and fast in detecting car accident. In additionally, a car accident detection system based on first-person videos,
considering of wide shooting scope of roadside cameras in which detected anomalies by predicting the future locations
CVIS, multi-scale feature fusion method and loss function of car participants and then monitoring the prediction accu-
with dynamic weights are utilized to improve performance racy and consistency metrics. These methods also have some
of detecting small objects. limitations because of low penetration of vehicular intelligent
The rest of this paper is organized as follows: Section 2 devices and shielding effects between vehicles.
gives an overviews of related work. We present the details of There are also some other methods which use roadside
our proposed method in Section 3. The performance evalua- devices instead of vehicular equipments to obtain and pro-
tion is discussed in Section 4. Finally, Section 5 conclude this cess video. [27] proposed a novel accident detection system
paper. at intersection, which composed background images from
image sequence and detected accidents by using Hidden
II. RELATED WORK Markov Model. [28] outlined a novel method for modeling
The car accident detection and notification method is a of interaction among multiple moving objects, and used the
challenging issue and has attracted a lot of attention from Motion Interaction Field to detect and localize car accidents.
researchers. They have proposed and applied various car ac- [29] proposed a novel approach for automatic road accident
cident detection methods. In generally, car accident detection detection, which was based on detecting damaged vehicles
methods are mainly divided into the following two kinds: from footage received from surveillance cameras installed
vehicle running condition-based and accident video features- in roads. In this method, Histogram of gradients (HOG)
based. and Gray level co-occurrence matrix features were used to
train support vector machines. [30] presented a novel dataset
A. METHOD BASED ON VEHICLE RUNNING CONDITION for car accidents analysis based on traffic Closed-Circuit
When an accident occurs, the motion state of the vehicle Television (CCTV) footage, and combined Faster Regions-
will change dramatically. Therefore, many researchers pro- Convolutional Neural Network (R-CNN) and Context Min-
posed the accident detection method by monitoring motion ing to detect and predict car accidents. The method in [30]
parameters, such as acceleration, velocity and so on. [22] achieved 1.68 seconds in terms of Time-To-Accident mea-
used On Board Diagnosis (OBD) system to monitor speed sure with an Average Precision of 47.25%. [8] proposed a
and engine status to detect a crash, and utilized smart-phone novel framework for automatic car accident detection, which
to report the accident by Wi-Fi or cellular network. [23] learned feature representation from the spatio-temporal vol-
developed an accident detection and reporting system using umes of raw pixel intensity instead of traditional hand-crafted
GPS, GPRS, and GSM. The speed of vehicle obtained from features. The experiments of method in [8] demonstrated it
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

can detect on average 77.5% accidents correctly with 22.5% 1) Data collection and annotation
false alarms. There are two major challenges in collecting car accidents
Compared with the methods based on vehicle running data:(1) Access: access to roadside traffic cameras data is of-
condition, these methods improve the detection accuracy and ten limited. In addition, the accident data from transportation
some of them even can predict accidents about 2 seconds be- administration is often not available for public uses because
fore they occur. To some extent, these methods are significant of many legal reasons. (2) Abnormality: car accidents are rare
in decreasing the accident rate and improving traffic safety. in the road compared with normal traffic conditions. In this
However, the detection accuracy of these methods is low and work, we try to draw support from video sharing websites to
the error rate is high, and the wrong accident information will search the videos and images including car accidents, such
have a great impact on the normal traffic flow. Concerning as news report and documentary. In order to improve the
the core issue mentioned above, in order to avoid the draw- applicability of our proposed method to roadside edge device,
backs of vehicular cameras, our proposed method utilizes the we only pick out the videos and images captured from a
roadside intelligent edge devices to obtain traffic video and traffic CCTV footage.
process image. Moreover, for sake of improving the accuracy
of accident detection method based on intelligent roadside
devices, we establish the CAD-CVIS dataset based on video
sharing websites, which is consisted of various kinds of
accident types, weather conditions and accident locations.
Moreover, we develop the model YOLO-CA to improve the FIGURE 2. Data collection and annotation for the CAD-CVIS dataset.
reliability and real-time performance among different traffic
conditions by combining deep learning algorithms and MSFF Through the above steps, we obtain 633 car accidents
method. scenes, 3255 accident key frames and 225206 normal frames.
Moreover, the car accident scene only occupies a small
III. METHODS part of each accident frame. We utilize LabelImg [31] to
A. METHOD OVERVIEW annotate the location of the accident in each frame in detail
to enhance the accuracy of locating accident. The high ac-
curacy enables emergency message be sent to the vehicles
Intelligent that are in the same direction as accident more efficiently
roadside devices
CAD-CVIS
Data-driven
Rescue agencies and decrease the impact to the vehicles that are in the
YOLO-CA opposite direction. The whole steps of data collection and
Real-time
image
annotation are shown in Fig. 2. The CAD-CVIS dataset is
Accident?
Y
made available for research use through https://github.com/
DSRC/5G zzzzzzc/Car-accident-detection.

Car accident!
2) Statistics of the CAD-CVIS
Statistics of the CAD-CVIS dataset can be found in Fig. 3.It
can be found that the CAD-CVIS dataset includes various
types of car accidents, which can improve the adaptability of
our method to different conditions. According to the number
of vehicles in the accident, the CAD-CVIS dataset includes
FIGURE 1. The application scenario of the automatic car accident detection 323 Single Vehicle Accident frames, 2449 Double Vehi-
method based on CVIS.
cle Accidents frames and 483 Multiple Vehicle Accidents
frames. Moreover, the CAD-CVIS dataset covers a variety
The Fig. 1 shows the application principle of our proposed of weather conditions, such as 2769 accident frames under
car accident detection method based CVIS. Firstly, the car ac- sunny condition, 268 frames under foggy condition, 52 acci-
cident detection application program with YOLO-CA model dent frames under rainy condition and 166 accident frames
is deployed on the edge server, which is developed based on under snowy condition. Besides, there are 2588 frames of
CAD-CVIS and deep learning algorithms. Then edge server accidents in the daytime and 667 accident frames at night.
receives and processes the real-time image captured by road- In addition, the CAD-CVIS dataset contains 2281 frames
side cameras. Finally, the roadside communication unit will of accidents occurring at the intersection, 596 frames in the
broadcast the accident emergency messages to the relevant urban road, 189 frames in the expressway and 189 frames in
vehicles and rescue agencies by DSRC and 5G networks. In the highway.
the rest of this section, we will present the details of CAD- Comparison between CAD-CVIS and related datasets can
CVIS and YOLO-CA model. be found in Table. 1. The A in Table. 1 responses that there
is annotation of car accident in the dataset. R responses that
B. CAD-CVIS the videos and frames captured from the roadside CCTV
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

such as YOLO [21](You Only Look Once) and SSD (Single

2400
3000
Shot MultiBox Detector) [38]. These algorithms implement
Number of traffic accident frames

Number of traffic accident frames

2700
2100
2400 location and classification by one CNN, which can provide
1800

1500
2100

1800
end to end detection service. Because of eliminating the
1200
1500 process of selecting the proposal regions, these algorithms
900
1200

900
are very fast and still has guaranteeing accuracy. Considering
600
600 that accident detection requires high real-time performance,
300
300

0 0
we design the deep neural network based on one-stage model
Single Double Multiple Sunny Foggy Rainy Snowy
Number of vehicles in the accident Weather condition YOLO [21].
(a) (b)
1) Network Design
3000
2400 YOLO utilizes its particular CNN to complete classification
Number of traffic accident frames

Number of traffic accident frames

2700

2400
2100
and location of multiple objects in an image at one time. In
1800
2100

1800 1500
the training process of YOLO, each image is divided into
1500
1200 S × S grids. If the center of an object falls into a grid cell,
1200

900
900
that grid cell is responsible for detecting that object [39].
600
600

300
300
This design can improve the detection speed dramatically
0
Daytime Night
0
Intersection Urban road Expressway Highway
and the detection accuracy with reference to global features.
Time of the accident Location of the accident
However, it also will cause serious detection error when there
(c) (d) are more than one objects in one grids. Roadside cameras
FIGURE 3. Number of accident frames in CAD-CVIS categorized by different have a wide scope of shooting, the accident area may be
indexes. (a) Accident Type (b) Weather condition (c) Accident time (d) Accident
location small in the image. Inspired of the multi-scale feature fusion
(MSFF) network, in order to improve the performance of
TABLE 1. Comparison between CAD-CVIS and related datasets model to detect small objects, we utilize 24 layers to achieve
image upsampling and obtain two different dimensional out-
Dataset name Scenes Frames or Duration A R M put tensors. This new car accident detection model is called
UCSD Ped2 77 1636 frames × X × as YOLO-CA, and the network structure diagram of YOLO-
CUHK Avenue 47 3820 frames × × X
DAD 620 2.4 hours X × X CA is shown as Fig. 4.
CADP 1416 5.2 hours × X X As shown in Fig. 4, YOLO-CA is composed of 228 neural
CAD-CVIS 632 3255+225206 frames X X X network layers, and the number of each kind of layer is
shown in Table. 2. These layers constitute many kinds of
basic components of YOLO-CA network, such as DBL and
footage. M responses that there are multiple road conditions ResN. The DBL is the minimum components of YOLO-
in dataset. Compared with CUHK Avenue [32], UCSD Ped2 CA network, which is composed of Convolution layer, Batch
[33] and DAD [25], CAD-CVIS contains more car accident Normalization layer and Leaky ReLU layer. ResN consists of
scenes, which can improve the adaptability of model based Zero Padding layer, DBL and N Resblock_units [40], which
on CAD-CVIS. Moreover, the frames of CAD-CVIS are all is designed to avoid neural network degradation caused by
captured from roadside CCTV footage, which is more suit- increased depth. Ups in Fig. 4 is upsampling layer, which is
able for the accident detection methods based on intelligent utilized to improve the performance of YOLO-CA to detect
roadside devices in CVIS. small objects. Concat is concatenate layer, which is used to
concatenate the layer in Darknet-53 and upsampling layer.
C. OUR PROPOSED DEEP NEURAL NETWORK MODEL
TABLE 2. Composition of YOLO-CA Network
In the task of car accident detection, we must not only
judge whether there is a car accident in the image, but
Layer name Number
also accurately locate the car accident. That’s because the Input 1
accurate location guarantees that the RSU can broadcast the Convolution 65
emergency message to the vehicles affected by the accident. Batch Normalization 65
Leaky ReLU 65
The classification and location algorithms can be divided Zero Padding 5
into two kinds:(1) Two stage model, such as R-CNN [34], Add 23
Fast R-CNN [35], Faster R-CNN [36] and Faster R-CNN Upsampling 1
Concatenate 1
with FPN [37]. These algorithms utilize selective research Total 228
and Region Proposal Network (RPN) to select about 2000
proposal regions in the image, and then detection objects
by the features of these regions extracted by CNN. These 2) Detection principle
region-based models locate objects accurately, but extracting Fig. 5 shows the detection principle of YOLO-CA, which
proposals take a great deal of time. (2) One stage model, includes extracting feature map and predicting bounding box.
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

DBL*5

DBL DBL Conv

Darknet-53 Output1
13x13x18
DBL Res1 Res2 Res8 Res8 Res4

Input DBL*5
416x416x3 DBL UpS
Concat
DBL DBL Conv
Output2
26x26x18

Darkmetconv2D_BN_Leaky Resblock_body Resblock_unit

Leaky Zero Res Res
DBL Conv BN ResN DBL DBL DBL Add
ReLU Padding Unit Unit
Res_Unit*N

FIGURE 4. The network structure of YOLO-CA.

will be calculated, which is utilized to adjust the weights of

13x13 grid
13x13x18 parameters. In the calculation of loss, let the ground truth
of an object is x∗ , y ∗ , w∗ , h∗ , CS ∗ , p∗ . S × S is the number
1
2
3

of cells in grid, and B is the number of predicted bounding

BBox1 BBox2 BBox3 boxes of each grid cell. For each grid cell, the P r(Objects)
equals 1 when the cell contains center of object, whereas it
equals 0 when there is not center of object in the cell. For each
26x26x18
12
3
image, the loss of YOLO-CA is divided into the following
416x416x3 four parts:

26x26 grid BBox1 BBox2 BBox3

• Loss of (x, y), which is calculated by (1). Where BCL

FIGURE 5. The detection principle of YOLO-CA. is binary cross entropy loss function, and the areaT ure
is defined as w∗ ∗ h∗ .

As shown in Fig. 5, YOLO-CA divides the input image into

13 × 13 grid and 26 × 26 grid. The first grid is responsible for
detecting the large objects, whereas the second grid makes up
for the inaccuracy of small target detection in the first grid. S×S
XX B
The feature extraction networks corresponding to these two Lossxy = P r (Objects) (2 − areaT ureij )
grids are different, but the detection models of the objects i=1 j=1
is similar. For ease of presentation, we regard the first grid ∗ [BCL (xij ) + BCL (yij )]
as example to explain the training steps of YOLO-CA. The BCL (xij ) = x∗ij log xij + 1 − x∗ij log (1 − xij )


center of car accident region falls into the grid cell (7, 5), so ∗ ∗


BCL (yij ) = yij log yij + 1 − yij log (1 − yij )
this cell is responsible for detecting this car accident in the
whole training process. Then the cell (7, 5) will predict three (1)
bounding boxes, and each boxes includes six parameters:
x, y, w, h, CS, p. The (x, y) is the center point of the bound-
ing box, and the (w, h) is the ratio of width and height of
the bounding box to the whole image. The CS is confidence
score of bounding box, which represents how confident the • Loss of (w, h), which is calculated by (2). Where
model is that the bounding box contains an object and how SD is square difference function. Especially, the (2 −
accurate it thinks the box is that it predicts. Lastly, each areaT ureij ) in (1) and (2) is utilized to increase the
bounding box will predict class probability of car accident error punishment of small objects. Because that the
p. same errors of x, y, w, h cause more serious impact on
After the training of a batch of images, the loss of model the detection effect of small object than that of large
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

b
1X
Loss = Loss_imgk (7)
Intersection b
k=1

where the b in (7) is the size of batch.

IoU tp =
IV. EXPERIMENT
Union In this section, we evaluate our proposed model YOLO-CA
on the CAD-CVIS dataset. First, we give the training results
of YOLO-CA, which include the change process of several
performance indexes. Then, we show the results of some
FIGURE 6. The definition of IoU. comparative experiments between YOLO-CA and other de-
tection models. Finally, the visual results are demonstrated
among various types and scales of car accident objects.
object.
S×S
XX B A. IMPLEMENTATION DETAILS
Losswh = P r (Objects) (2 − areaT ureij ) We implement our model in TensorFlow [41] under the op-
i=1 j=1 erating system Ubuntu 18.04 and perform experiments on a
1 system with Nvidia Titan Xp GPU. We divide the CAD-CVIS
∗ [SD (wij ) + SD (hij )]
2 dataset into three parts: (1) Training set (80%), which is used
∗ 2


SD (wij ) = wij − wij to train the parameter weight of network. (2) Validation set

2 (5%), which is utilized to adjust hyperparameters, such as
SD (hij ) = hij − h∗ij
learning rate and drop out rate. (3) Test set (15%), which is
(2) used to evaluate the performance of different algorithms for
• Loss of CS, which is calculated by (3). The loss of CS detecting car accident. In additionally, each part of dataset
can be divided into two parts: the confidence loss of contains all types of accident in Fig. 3. The batch size is set to
foreground and confidence loss of background. 64, and the models are trained for up to 30000 iterations. The
initial learning rate is set to 0.001, and updating with iteration
S×S B
XX parameter of 0.1/10000 iterations. The SGD optimizer with a
LossCS = P r (Objects) ∗ BCL (CSij ) momentum of 0.9 is utilized to adjust parameters of network.
i=1 j=1
Moreover, we use a weight decay of 0.0005 to prevent model
+ (1 − P r (Objects)) ∗ BCL (CSij ) overfitting.
∗
BCL (CSij ) = CSij log CSij
∗ B. RESULTS AND ANALYSIS


+ 1 − CSij log (1 − CSij )
1) Training results of YOLO-CA
(3)
Fig. 7 shows the training results of YOLO-CA, including
Where CS ∗ is defined by (4). In additionally, the IoUpt the changes of precision, recall, IoU and loss of each batch
is defined in Fig. 6, which equals the intersection over in iteration process. In the training process of YOLO-CA,
union (IoU) between the predicted bounding box and the we regard the prediction result with IoU over 0.5 and right
ground truth. classification as true result, and other predictions are all false
results. As shown in Table. 3, the prediction results can be
CS ∗ = Pr(Objects) ∗ IoUpt (4)
divided into four parts: (1) TP: Truth Positive. (2) FP: False
• Loss of p, which is calculated by (5) Positive. (3) FN: False Negative. (4) TN: True Negative. The
S×S B
precision is defined as precision = T PT+F P
P and recall is
TP
defined as recall = T P +F N .
XX
LossCS = P r (Objects) BCL (pij )
i=1 j=1
TABLE 3. Distribution map of prediction results
p∗ij log pij + 1 − p∗ij log (1 − pij )


BCL (pij ) =
(5) Ground truth
Positive Negative
For each image in training set, the total loss is defined as (6). Prediction Positive TP FP
Especially, because that the multi-scale feature fusion is used Result Negative FN TN
in YOLO-CA, the loss is the sum of conditions under S = 13
and S = 26. In additionally, the loss of each batch of images As shown in Fig. 7a, with the increasing of iterations, the
is defined as (7). precision of YOLO-CA is increasing gradually and converge
over 90%. Moreover, recall eventually converges to more
Loss_img = Lossxy + Losswh + LossCS + Lossp (6) than 95%. In terms of locating performance of YOLO-CA
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

defined as the average number of frames that can be

1 1
detected in one second. This indexes evaluate real-time
0.8 0.8 performance of detection models.
Precision

0.6 0.6

Recall
0.4 0.4

0.2 0.2

1
0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Batches 104 Batches 104
0.9

(a) (b)
0.8

1 2

Precision
0.7

0.8 1.6
0.6

0.6 1.2
Fast R-CNN (AP=77.65%)
Loss

0.5
IoU

Faster R-CNN (AP=82.02%)

0.4 0.8 Faster R-CNN with FPN (AP=90.66%)
0.4 SSD (AP=83.40%)
YOLOv3 without MSFF (AP=83.16%)
0.2 0.4
YOLOv3 (AP=86.20%)
0.3
YOLO-CA (AP=90.02%)
0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
4 4 0.2
Batches 10 Batches 10
0 0.2 0.4 0.6 0.8 1
Recall
(c) (d)
FIGURE 7. The training results of YOLO-CA. (a) Precision (b) Recall (c) IoU (a)
(d) Loss

in training set, IoU finally stabilizes above 0.8. The Fig. 7d 0.8

shows the decreasing process of loss of YOLO-CA in (7),

and the final convergence of loss is less than 0.2. 0.6
Average IoU

2) Comparative experiments and visual results 0.4

The comparative experiments are conducted for comparing

seven detection models: (1) One-stage models: SSD, our 0.2

proposed YOLO-CA, traditional YOLO-v3 and YOLO-v3

without MSFF (Multi-Scale Feature Fusion). (2) Two-stage 0
Fast
R-CNN
Faster
R-CNN
Faster
R-CNN
SSD YOLOv3
without
YOLOv3 YOLO-CA

with FPN MSFF

models: Fast R-CNN, Faster R-CNN and Faster R-CNN with
FPN. In order to comparatively demonstrate the validation
of YOLO-CA as well as confirm its strength in terms of the (b)
comprehensive performance on the accuracy and real-time, FIGURE 8. The AP and IoU results of different models. (a) Precision-Recall
the following indexes are selected for comparison among the curve (b) Average IoU
seven models:
• Average Precision (AP) that is defined as the average The Fig. 8a shows the Precision-Recall curve of detection
value of precision under different recall, which can be models among test set. It can be found that Faster R-CNN
changed by adjusting threshold of classification confi- with FPN and our proposed YOLO-CA have obvious advan-
dence. AP index evaluate the accuracy performance of tages in accuracy performance than the other models. In addi-
detection models. The average precision can be calcu- tionally, our-proposed YOLO-CA can achieve 90.02% of AP,
lated by (8). which is slightly lower than that of Faster R-CNN with FPN
Z 1 (90.66%) and higher than those of Fast R-CNN (77.66%),
AP = precision(r) (8) Faster R-CNN (82.02%), SSD (83.40%), YOLOv3 without
0 MSFF (83.16%) and YOLOv3 (86.20%). Average IoU is a
where r is recall. vital important index to evaluate locating performance of
• Average Intersect over Union (Average IoU) that is detection models. Moreover, accurate location is critical to
utilized to evaluate the object locating performance of car accident detection and notification, and higher locating
detection models. The Average IoU is the average value performance can improve the safety of the vehicles around
of IoUs between every prediction bounding box and accident. As shown in Fig. 8b, YOLO-CA can achieve about
corresponding ground truth. 0.73 of Average IoU, which is lower than that of Faster R-
• Frames Per Second (FPS). Inference time is defined as CNN with FPN (0.75) and higher than those of Fast R-CNN
the average time cost of detecting a frame among test (0.58), Faster R-CNN (0.66), SSD (0.69), YOLOv3 without
set. FPS is the reciprocal of inference time, which is MSFF (0.65) and YOLOv3 (0.71).
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

TABLE 4. AP and IoU results of different models among different scales of object

AP(%) Average IoU

Method Large Medium Small Large Medium Small
Fast R-CNN 82.11 80.3 49.20 0.68 0.63 0.46
Faster R-CNN 89.66 86.79 56.79 0.71 0.69 0.51
Faster R-CNN with FPN 93.07 92.15 78.63 0.74 0.72 0.68
SSD 91.37 89.65 60.59 0.73 0.71 0.58
YOLOv3 without MSFF 90.18 90.40 58.89 0.72 0.68 0.50
YOLOv3 90.45 91.17 67.78 0.74 0.73 0.60
YOLO-CA 93.87 91.51 76.51 0.76 0.74 0.64

In order to compare and analysis the performance of mod- loss among different scales of objects. This process increases
els in details, the objects of test set is divided into three parts the error punishment of small objects, because that the
according to different scales of objects:(1) Large: the area of same errors of x, y, w, h cause more serious impact on the
object is larger than one tenth of image size. (2) Medium: detection effect of the small object than that of the large
the area of object is over the interval [1/100, 1/10] of image object. Consequently, YOLO-CA has obvious advantages in
size. (3) Small: the area of object is less than one-hundredth AP and Average IoU of small objects than YOLOv3. The
of image size. MSFF processes of Faster R-CNN with FPN and YOLO-
CA are similar, feature pyramid networks is used to extract
The Table. 4 shows the AP and IoU results of the seven
feature maps of different scales and fuse these maps to obtain
models among different scales of object. We can intuitively
features with high-semantic and high-resolution. Faster R-
see that the scales of objects significantly affect the accuracy
CNN utilizes RPN to select about 20000 proposal regions,
and locating performance of detection models. It can be
whereas there are only 13∗13∗3+26∗26∗3 = 2535 candidate
found that our proposed YOLO-CA has obvious advantages
bounding boxes in YOLO-CA. This difference results in
in AP and Average IoU than Fast R-CNN, Faster R-CNN
Faster R-CNN has slight advantages in accuracy performance
and YOLOv3 without MSFF, especially among small scale
than YOLO-CA, but also causes serious disadvantages in
of objects. There is not MSFF process in the above three
real-time performance.
models, which results in that they detection the objects only
rely on the top-level features. However, although there is
rich semantic information in top-level features, the location
30
information of objects is rough, which does not benefit to
27
locate the bounding box of objects correctly. On the contrary,
24
there is little semantic information in low-level features with
high resolution, but the location information of objects is 21

accurate. For small scales of objects, they make up a small 18

FPS

proportion of the whole frame, and their location information 15

is easily lost through multiple convolution processes. YOLO- 12

CA utilizes MSFF to combine top-level features and low- 9

level features, and then makes a prediction in each fused 6

feature layer. This process reserves the rich semantic in- 3

formation and accurate location information simultaneously, 0

Fast
R-CNN
Faster
R-CNN
Faster
R-CNN
SSD YOLOv3
without
YOLOv3 YOLO-CA

so YOLO-CA has better performance in AP and Average with FPN MSFF

IoU than Fast R-CNN, Faster R-CNN and YOLOv3 with-

out MSFF. For SSD, it uses pyramidal feature hierarchy to
obtain multi-scale feature maps. But to avoid using low- FIGURE 9. The FPS of different models.

level features SSD foregoes reusing already computed layers

and instead builds the pyramid starting from high up in the Fig. 9 shows the FPS results of different models among
network and then by adding several new layers. So SSD test set. It can be found that the FPS of one-stage models
misses the opportunity to reuse the higher-resolution maps is obviously higher than that of two-stage models. This low
of feature hierarchy, which are vital important for detecting performance of the two-stage models results from a great deal
small objects [37]. Moreover, the performance of backbone of time cost of selecting proposal regions.
of YOLO-CA (Darknet53) is better than that of SSD (VGG- Fast R-CNN utilizes time-consuming selective research al-
16) because of using residual networks to avoid degradation gorithm to select proposal regions based on color and texture
problem of deep neural network. Therefore, YOLO-CA can features, which results in that Fast R-CNN only achieves 0.4
achieve better results of AP and Average IoU than SSD. of FPS. Faster R-CNN uses the RPN that share convolutional
Compared with YOLOv3, YOLO-CA utilizes loss function layers with state-of-the-art object detection networks instead
with dynamic weights to balance the influence of location of selective research to generate proposals. Benefiting from
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

RPN, Faster R-CNN achieve about 3.5 of FPS among test set 4) Comparison with other car accident detection methods
(Faster R-CNN:3.5, Faster R-CNN with FPN:3.6). Although other car accident detection methods utilize a small
Although Faster R-CNN obtains significantly improve- private collection of datasets and do not make them public
ment of real-time performance compared with Fast R-CNN, so comparing them may not be fair at this stage. But still,
there is still a big gap with one-stage models. That is be- we list the performance achieved by these methods on their
cause one-stage models abandon the process of selecting individual datasets. ARRS [3] achieve about 63% AP with
proposal regions and utilize one CNN to implement location 6% false alarms. The method of [42] achieve 89.50% AP.
and classification of objects. As shown in Fig. 9, SSD can DSA-RNN [25] achieve about 80% recall and 56.14% AP.
achieve 15.6 of FPS among test set. The other three models The method in [30] achieve about 47.25% AP. The method
based on YOLO utilize the backbone Darknet-53 instead of of [8] achieve 77.5% AP and 22.5% false alarms. Moreover,
VGG-16 in SSD, and computation of the former network the number of accident scenes of the datasets utilized in these
is significantly less than the latter because of using the methods is limited, which will result in poor adaptability for
residual networks. Therefore, the real-time performance of new scenarios.
SSD is lower than YOLO-based models in our experiments.
In additionally, our proposed YOLO-CA simplifies the MSFF V. CONCLUSION
networks of YOLOv3. So YOLO-CA can achieve 21.7 of In this paper, we have proposed an automatic car accident
FPS, which is higher than that of YOLOv3 (about 19.1). Be- detection method based on CVIS. First of all, we present the
cause of lacking MSFF process in YOLOv3 without MSFF, application principles of our proposed method in the CVIS.
it has better real-time performance (about 23.6 of FPS) than Secondly, we build a novel image dataset CAD-CVIS, which
YOLO-CA, but this lacking results in serious performance is more suitable for car accident detection method based on
penalties of AP. intelligent roadside devices in CVIS. Then we develop the car
Fig. 10 show some visual results of the seven models accident detection model YOLO-CA based on CAD-CVIS
among different scales of objects. It can be found that there and deep learning algorithms. In the model, we combine the
is a false positive in the large objects detection results of Fast multi-scale feature fusion and loss function with dynamic
R-CNN, but the other six models all have high accuracy and weights to improve real-time and accuracy of YOLO-CA.
locating performance in large objects in Fig. 10. However, Finally, we show the simulation experiments results of our
the locating performance of Fast R-CNN, Faster R-CNN, method, which demonstrates our proposed methods can de-
SSD, and YOLOv3 without MSFF decrease significantly in tect car accident in 0.0461 seconds with 90.02% AP. More-
medium object frame (1), and the prediction bounding box over, the comparative experiments results show that YOLO-
cannot fitting out the contour of car accident. Moreover, Fast CA has comprehensive performance advantages of detecting
R-CNN, SSD, and YOLO-without MSFF cannot detect the car accident than other detection models, in terms of accuracy
car accident in small object frame (1). In additionally, except and real-time.
for Faster R-CNN with FPN and YOLO-CA, other models
have serious location error in small object frame (3). REFERENCES
[1] WHO, “Global status report on road safety 2018,” https://www.who.int/
3) Comparison of comprehensive performance and violence_injury_prevention/road_safety_status/2018/en/.
[2] H. L. Wang and M. A. Jia-Liang, “A design of smart car accident rescue
practicality system combined with wechat platform,” Journal of Transportation Engi-
As analyzed above, it can be found that our proposed YOLO- neering, 2017.
CA has performance advantages of detecting car accident [3] Y. Ki and D. Lee, “A traffic accident recording and reporting model at
intersections,” IEEE Transactions on Intelligent Transportation Systems,
than Fast R-CNN, Faster R-CNN, SSD, and YOLOv3 in vol. 8, no. 2, pp. 188–194, June 2007.
terms of accuracy, locating and real-time performance. For [4] W. Hao and J. Daniel, “Motor vehicle driver injury severity study under
YOLOv3 without MSFF, the FPS of it (23.6) is higher than various traffic control at highway-rail grade crossings in the united states,”
Journal of Safety Research, vol. 51, pp. 41–48, 2014.
that of YOLO-CA (21.7), and this difference is acceptable in [5] J. White, C. Thompson, H. Turner, B. Dougherty, and D. C. Schmidt,
the practical application of detecting car accident. However, “Wreckwatch: Automatic traffic accident detection and notification with
the AP of YOLO-CA is significantly higher than that of smartphones,” Mobile Networks and Applications, vol. 16, no. 3, pp. 285–
303, 2011.
YOLOv3 without MSFF, especially for small scales of object [6] S. Sadek, A. Al-Hamadi, B. Michaelis, and U. Sayed, “Real-time auto-
(76.51% vs 58.89%). Compared with Faster R-CNN with matic traffic accident recognition using hfg,” in International Conference
FPN, YOLO-CA can approach the AP of it (90.66% vs on Pattern Recognition, 2010.
[7] A. Shaik, N. Bowen, J. Bole, G. Kunzi, D. Bruce, A. Abdelgawad, and
90.03%) with an obvious speed advantage. Faster R-CNN K. Yelamarthi, “Smart car: An iot based accident detection system,” in
cost about 277ms on average to detect one frame, whereas 2018 IEEE Global Conference on Internet of Things (GCIoT). IEEE,
YOLO-CA only need 46 ms, which illustrates the speed 2018, pp. 1–5.
[8] D. Singh and C. K. Mohan, “Deep spatio-temporal representation for
of YOLO-CA is about 6x faster than Faster R-CNN with detection of road accidents using stacked autoencoder,” IEEE Transactions
FPN. Car accident detection in CVIS requires high real-time on Intelligent Transportation Systems, vol. 20, no. 3, pp. 879–887, March
performance because of the high dynamics of vehicles. To 2019.
[9] M. Zheng, T. Li, R. Zhu, J. Chen, Z. Ma, M. Tang, Z. Cui, and Z. Wang,
summarize, our proposed YOLO-CA have higher practicality “Traffic accident’s severity prediction: A deep-learning approach-based
and comprehensive performance on accuracy and real-time. cnn network,” IEEE Access, vol. 7, pp. 39 897–39 910, 2019.

VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

Faster R-CNN YOLOv3

Fast R-CNN Faster R-CNN SSD
With FPN Without MSFF YOLOv3 YOLO-CA

(1)

Large (2)

(3)

(1)

Mediun (2)

(3)

(1)

Small (2)

(3)

FIGURE 10. Some visual results of the seven models among different scales of objects.

[10] L. Zheng, Z. Peng, J. Yan, and W. Han, “An online learning and unsuper- [16] S. Ramos, S. Gehrig, P. Pinggera, U. Franke, and C. Rother, “Detecting
vised traffic anomaly detection system,” Advanced Science Letters, vol. 7, unexpected obstacles for self-driving cars: Fusing deep learning and
no. 1, pp. 449–455, 2012. geometric modeling,” in 2017 IEEE Intelligent Vehicles Symposium (IV).
[11] F. Yang, S. Wang, J. Li, Z. Liu, and Q. Sun, “An overview of internet of IEEE, 2017, pp. 1025–1032.
vehicles,” China Communications, vol. 11, no. 10, pp. 1–15, Oct 2014. [17] T. Qu, Q. Zhang, and S. Sun, “Vehicle detection from high-resolution
[12] C. Ma, W. Hao, A. Wang, and H. Zhao, “Developing a coordinated aerial images using spatial pyramid pooling-based deep convolutional
signal control system for urban ring road under the vehicle-infrastructure neural networks,” Multimedia Tools and Applications, vol. 76, no. 20, pp.
connected environment,” IEEE Access, vol. 6, pp. 52 471–52 478, 2018. 21 651–21 663, 2017.
[13] S. Zhang, J. Chen, F. Lyu, N. Cheng, W. Shi, and X. Shen, “Vehicular [18] D. Dooley, B. McGinley, C. Hughes, L. Kilmartin, E. Jones, and
communication networks in the automated driving era,” IEEE Communi- M. Glavin, “A blind-zone detection method using a rear-mounted fisheye
cations Magazine, vol. 56, no. 9, pp. 26–32, 2018. camera with combination of vehicle detection methods,” IEEE Transac-
[14] Y. Wang, D. Zhang, Y. Liu, B. Dai, and L. H. Lee, “Enhancing transporta- tions on Intelligent Transportation Systems, vol. 17, no. 1, pp. 264–278,
tion systems via deep learning: A survey,” Transportation research part C: Jan 2016.
emerging technologies, 2018. [19] X. Changzhen, W. Cong, M. Weixin, and S. Yanmei, “A traffic sign
[15] G. Wu, F. Chen, X. Pan, M. Xu, and X. Zhu, “Using the visual intervention detection algorithm based on deep convolutional neural network,” in 2016
influence of pavement markings for rutting mitigation–part i: preliminary IEEE International Conference on Signal and Image Processing (ICSIP),
experiments and field tests,” International Journal of Pavement Engineer- Aug 2016, pp. 676–679.
ing, vol. 20, no. 6, pp. 734–746, 2019. [20] S. Zhang, C. Bauckhage, and A. B. Cremers, “Efficient pedestrian detec-

10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2939532, IEEE Access

Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

tion via rectangular features based on a statistical shape model,” IEEE

Transactions on Intelligent Transportation Systems, vol. 16, no. 2, pp. 763–
775, April 2015.
[21] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv:
Computer Vision and Pattern Recognition, 2018.
[22] J. Zaldivar, C. T. Calafate, J. C. Cano, and P. Manzoni, “Providing accident
detection in vehicular networks through OBD-II devices and android-
based smartphones,” Proceedings - Conference on Local Computer Net-
works, LCN, no. April 2014, pp. 813–819, 2011.
[23] M. S. Amin, J. Jalil, and M. B. I. Reaz, “Accident detection and reporting
system using gps, gprs and gsm technology,” in International Conference
on Informatics, 2012.
[24] M. Fogue, P. Garrido, F. J. Martinez, J. C. Cano, C. T. Calafate, and P. Man-
zoni, “Automatic accident detection: Assistance through communication
technologies and vehicles,” IEEE Vehicular Technology Magazine, vol. 7,
no. 3, pp. 90–100, 2012.
[25] F.-H. Chan, Y.-T. Chen, Y. Xiang, and M. Sun, “Anticipating accidents
in dashcam videos,” in Computer Vision – ACCV 2016. Springer
International Publishing, 2017, pp. 136–153.
[26] Y. Yao, M. Xu, Y. Wang, D. J. Crandall, and E. M. Atkins, “Unsupervised
traffic accident detection in first-person videos.” arXiv: Computer Vision
and Pattern Recognition, 2019.
[27] S. Kamijo, Y. Matsushita, K. Ikeuchi, and M. Sakauchi, “Traffic Mon-
itoring and Accident Detection at Intersections,” IEEE Transactions on
Intelligent Transportation Systems, vol. 1, no. 2, pp. 108–117, 2000.
[28] K. Yun, H. Jeong, K. M. Yi, S. W. Kim, and J. Y. Choi, “Motion interaction
field for accident detection in traffic surveillance video,” Proceedings -
International Conference on Pattern Recognition, pp. 3062–3067, 2014.
[29] V. Ravindran, L. Viswanathan, and S. Rangaswamy, “A Novel Approach to
Automatic Road-Accident Detection using Machine Vision Techniques,”
International Journal of Advanced Computer Science and Applications,
vol. 7, no. 11, pp. 235–242, 2016.
[30] A. Shah, J. Lamare, T. Nguyenanh, and A. G. Hauptmann, “Cadp: A novel
dataset for cctv traffic camera based accident analysis,” advanced video
and signal based surveillance, pp. 1–9, 2018.
[31] Tzutalin, “Labelimg,” Git code, 2015,
https://github.com/tzutalin/labelImg.
[32] C. Lu, J. Shi, and J. Jia, “Abnormal event detection at 150 FPS in MAT-
LAB,” Proceedings of the IEEE International Conference on Computer
Vision, pp. 2720–2727, 2013.
[33] V. Mahadevan, W. Li, V. Bhalodia, and N. Vasconcelos, “Anomaly
detection in crowded scenes,” 2010 IEEE Computer Society Conference
on Computer Vision and Pattern Recognition, pp. 1975–1981, 2010.
[Online]. Available: http://ieeexplore.ieee.org/document/5539872/
[34] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies
for accurate object detection and semantic segmentation,” in The IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), June
2014.
[35] R. Girshick, “Fast r-cnn,” Computer Science, 2015.
[36] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: towards real-
time object detection with region proposal networks,” in International
Conference on Neural Information Processing Systems, 2015.
[37] T.-Y. Lin, P. Dollár, R. B. Girshick, K. He, B. Hariharan, and S. J. Belongie,
“Feature pyramid networks for object detection,” 2017 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), pp. 936–944, 2016.
[38] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Y. Fu, and A. C.
Berg, “Ssd: Single shot multibox detector,” in European Conference on
Computer Vision, 2016.
[39] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi, “You only look
once: Unified, real-time object detection,” computer vision and pattern
recognition, pp. 779–788, 2016.
[40] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2016, pp. 770–778.
[41] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale ma-
chine learning on heterogeneous distributed systems,” arXiv: Distributed,
Parallel, and Cluster Computing, 2015.
[42] K. Yun, H. Jeong, K. M. Yi, S. W. Kim, and J. Y. Choi, “Motion interaction
field for accident detection in traffic surveillance video,” in 2014 22nd
International Conference on Pattern Recognition, Aug 2014, pp. 3062–
3067.

VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.