We introduce a challenging public dataset for radar detection (RadDet), comprising a large corpus of radar signals occupying a wideband spectrum across diverse radar density environments and signal-to-noise ratio settings. This repo contains the download links to our radar dataset and the associated conference paper "RadDet: A Wideband Dataset for Real-Time Radar Spectrum Detection". This work was accepted for publication at the 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2025) in Hyderabad, India.

You can access our preprint 📄 here: https://arxiv.org/abs/2501.10407

You can also watch our ICASSP 2025 presentation ▶️ here: https://youtu.be/H6LI_ZrdgeI

Huang, Z., Denman, S., Pemasiri, A., Martin, T., & Fookes, C. (2025). RadDet: A wideband dataset for real-time radar spectrum detection. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1–5. IEEE.

Our experiments considered several radar datasets:

1. RadDet-1T - this is our proposed dataset, it contains up to 9 signal targets per frame.
2. RadDet-9T - this is our proposed dataset, it contains at most 1 signal target per frame.
3. NIST-CBRS - this is a radar dataset created by NIST containing at most 1 signal target per frame. We adapted the original dataset and modified it in our experiments, the modified version of the dataset (NIST-CBRS) is provided here for reference. Please cite the original work by NIST if you wish to adapt their dataset in your research.

Each radar dataset contains max-hold spectrograms provided in three resolutions:

• 128 x 128 - 128 by 128 spectrograms, please refer to our paper for the specific t-f resolution.
• 256 x 256 - 256 by 256 spectrograms, please refer to our paper for the specific t-f resolution.
• 512 x 512 - 512 by 512 spectrograms, please refer to our paper for the specific t-f resolution.

RadDet introduces 11 radar classes, including 6 new LPI polyphase codes (P1, P2, P3, P4, Px, Zadoff-Chu) and a new wideband frequency-modulated continuous wave (FMCW), all coexisting across a 500 MHz band.

RadDet contains a total of 40,000 radar frames provided in three parts:

• Training set - contains 20,000 frames.
• Validation set - contains 14,000 frames.
• Test set - contains 6,000 frames.

We sample SNR from a uniform distribution to produce signal frames that fall within −20 and 20 dB at a resolution of 8 dB. This means that there are more than 6,500 unique signals genreated per SNR.

To investigate wideband spectrum detection in different scenarios, we provide RadDet in two different radar environments:

• Our sparse dataset (RadDet-1T) provides at most a single radar instance per frame whereby the probability of a radar being present in a scene is 50%.
• Our dense dataset (RadDet-9T) contains up to 9 radar instances per frame where the probability of background (noise-only) frames is 10%. RadDet-9T represents a conservative dense maritime radar environment.

Each dataset contains a configuration file called data.yaml provided in the standard YOLO-format. You can use this configuration file to create custom data modules for your project.

An example configuration file is shown below:

# Dataset root dir
path: ../Datasets/<DATASET_NAME>

# Train, validation, and test dataset paths (relative to path)
train: images/train
val: images/val
test: images/test

# Signal classes
names: 
  0: Rect
  1: Barker
  2: Frank
  3: P1
  4: P2
  5: P3
  6: P4
  7: Px
  8: ZadoffChu
  9: LFM
  10: FMCW

The bounding box annotations are provided as .txt files following the standard YOLO-format. Each .txt file corresponds to a unique .png file with the same name, i.e., train/000000000000.txt contains labels for train/000000000000.png.

An example .txt file containing 6 bounding boxes:

0 0.611153 0.540000 0.250000 0.039563
9 0.263525 0.460000 0.040000 0.116000
4 0.273151 0.780000 0.550000 0.041063
3 0.804995 0.620000 0.055000 0.041920
3 0.259591 0.660000 0.035000 0.041920

The download links for each radar dataset are provided below. The total size of the combined datasets is approximately 54 GB. We provide the data in three different resolutions in this release:

⚙️ Low resolution: 128 x 128

• RadDet-1T-128 - approx. 556 MB
• RadDet-9T-128 - approx. 577 MB
• NIST-CBRS-128 - approx. 1.2 GB

⚙️ Medium resolution: 256 x 256

• RadDet-1T-256 - approx. 2.4 GB
• RadDet-9T-256 - approx. 2.4 GB
• NIST-CBRS-256 - approx. 4.8 GB

⚙️ High resolution: 512 x 512

• RadDet-1T-512 - approx. 11.2 GB
• RadDet-9T-512 - approx. 11.6 GB
• NIST-CBRS-512 - approx. 18.9 GB

You can also download the original (unmodified) NIST dataset here. Please also cite the original work by NIST if you wish to use their dataset in your research.

💡 Please cite our conference paper if you find it helpful for your research. Cheers.

@inproceedings{huang2025raddet,
  title={RadDet: A Wideband Dataset for Real-Time Radar Spectrum Detection},
  author={Huang, Zi and Denman, Simon and Pemasiri, Akila and Martin, Terrence and Fookes, Clinton},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

💡 Our previous work may also be of interest to you:

@inproceedings{huang2024multi,
  title={Multi-Stage Learning for Radar Pulse Activity Segmentation},
  author={Huang, Zi and Pemasiri, Akila and Denman, Simon and Fookes, Clinton and Martin, Terrence},
  booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7340--7344},
  year={2024},
  organization={IEEE}
}