RADTrack

RADTrack is the modified version of the RADDet dataset which incorporates the Multi-Object Tracking (MOT) use case. RADTrack transforms the RADDet data into sequences showing different times and locations where the data was recorded. Labeled object IDs allow for tracking of these objects using Range-Azimuth-Doppler (RAD) data.

With the RADTrack dataset, you can:

• Train and evaluate object detection models on Range-Azimuth, Range-Doppler, and Cartesian Radar data
• Train and evaluate object classification models on Range-Azimuth, Range-Doppler, and Cartesian Radar data
• Train and evaluate object tracking models on Range-Azimuth, Range-Doppler, and Cartesian Radar data

Paper

• This dataset is introduced and used by RadarMOTR which is accepted in 2024 International Radar Conference (RADAR).
• Accepted Preprint
• IEEE Xplore Version

Radar(stationary) Dataset for Dynamic Road Users

Dataset link

• Nextcloud

Dataset format

The folder structure and additional metadata in RADTrack are designed to mirror the MOT20 format. This similarity enables seamless integration with standard evaluation tools like TrackEval, requiring only minor adjustments to the ground truth (GT) data format.

|-- {RADTrack ROOT}
|   |-- radtrack-train
|   |   |-- radtrack0001
|   |   |   |-- RAD
|   |   |   |   |-- 000001.npy
|   |   |   |   |-- 000002.npy
|   |   |   |   |-- ...
|   |   |   |-- stereo_image
|   |   |   |   |-- 000001.jpg
|   |   |   |   |-- 000002.jpg
|   |   |   |   |-- ...
|   |   |   |-- gt
|   |   |   |   |-- gt.json
|   |   |   |-- frame_mapping.txt
|   |   |   |-- seqinfo.ini
|   |   |-- ...
|   |-- radtrack-val
|   |   |-- ...
|   |-- radtrack-test
|   |   |-- ...
|   |-- seqmaps
|   |   |-- radtrack-train.txt
|   |   |-- radtrack-val.txt
|   |   |-- radtrack-test.txt
|   |-- sensors_para
|   |   |-- registration_matrix
|   |   |   |-- ...
|   |   |-- stereo_para
|   |   |   |-- ...
|   |   |-- radar_config.json

Dataset details

The RADTrack dataset comprises 10,158 frames, organized into 24 sequences. To ensure consistency, the same radar configuration throughout the entire data collection process was used. The specifics of the data capture settings are summarized below and can also be found in the sensors_para/radar_config.json file.

"designed_frequency":  76.8, // Hz
"config_frequency":    77, // Hz
"maximum_range":       50, // m
"range_size":          256,
"azimuth_size":        256,
"doppler_size":        64,
"range_resolution":    0.1953125, // m/bin
"angular_resolution":  0.006135923, // radian/bin
"velocity_resolution": 0.41968030701528203, // (m/s)/bin

The dataset consists of 6 classes and various input and ground truth formats. Below is a summary of the information stored in the dataset.

• RAD: 3D-FFT radar data stored as a matrix of complex64 numbers with dimensions (256, 256, 64) in NumPy format.
• stereo_image: Two rectified stereo images.
• gt: A JSON file containing a list of dictionaries. Each dictionary corresponds to a specific frame and includes a list of entries, where the number of entries is equal to the number of radar objects present in that frame. The dictionary has the following keys:
  • classes: A list of class labels. There are 6 classes: person, bicycle, car, motorcycle, bus, and truck.
  • boxes: A list of bounding box coordinates in the format [x_center, y_center, z_center, w, h, d], where x represents Range, y represents Angle, and z represents Doppler axis.
  • cart_boxes: A list of Cartesian bounding box coordinates in the format [y_center, x_center, h, w].
  • ids: A list of object IDs.
• sensors_para: Includes stereo_para for stereo depth estimation and registration_matrix for cross-sensor registration.
• seqmaps: A CSV file for each split, indicating which sequences are included.
• frame_mapping.txt: A CSV file containing a mapping of matching frames between the RADDet and RADTrack datasets for backwards compatibility.
• seqinfo.ini: A file containing meta information about each sequence, such as its length.

Note

The stereo_para includes left_maps.npy and right_maps.npy, which are derived from cv2.initUndistortRectifyMap(...) and contain maps in both x and y directions. All other matrices are derived from cv2.stereoRectify(...).

Dataset splits

The dataset is divided into three splits: 70% for training, 20% for validation, and 10% for testing. If you need to modify these splits, you can merge them by copying the sequence directories into the desired split folder and updating the corresponding seqmaps files.

Statistics

Statistics	Test	Train	Validation	Overall
Frames	585	7528	2045	10158
Objects	88	1313	295	1696
Frames with one object	104	1273	499	1876
Frames with two objects	166	1918	758	2842
Frames with three objects	116	2071	441	2628
Frames with four objects	87	1249	238	1574
Frames with five objects	85	694	71	850
Frames with six objects	26	252	35	313
Frames with seven or more objects	1	71	3	75

Classes	Test	Train	Validation	Overall
cars	64	926	203	1193
trucks	13	202	45	260
buses	0	10	1	11
bicycles	2	24	9	35
persons	9	153	36	198
motorcycles	0	4	1	5

Dataset license

The tools and code are licensed under MIT. The dataset is licensed under Creative Commons Attribution 4.0 International License.

RADTrack Labeling/Viewing Tool

The labeling tool for the annotation of the object IDs and class can be located in the labeling_tool folder. It also can be used as a dataset viewer.

Installation

Requirements:

• Dataset
• Anaconda or Miniconda used as virtual environment (Skip if python venv is used)

1. Clone repository

   git clone https://github.com/madeit07/RADTrack.git
   cd RADTrack/labeling_tool

2. Create virtual environment (e.g. using conda)

   conda create -n radtrack python=3.11
   conda activate radtrack

3. Install other requirements

   pip install -r requirements.txt

4. (Optional) If you want to use the scripts in the tools folder, install those requirements as well:

   pip install -r ../tools/requirements.txt
   conda install ffmpeg # Only for tools/visualize.py

Run Labeling Tool

Execute Python script in console:

python radtrack_labeler.py

Under File->New... create a new project.

After the project is opened, all RAD data is processed and cached so that scrubbing through the images is responsive. This is done only once and can a long time.

Default project

If you don't want to load the project every time you start the tool, you can set a default project which gets automatically loaded on startup. Create a new project and configure it. Name it Default and save it under labeling_tool\projects\.

Keybinds

Playback Controls:

Action	Keybind	Alternative Keybind
Next frame	D	→
Previous frame	A	←
Jump to first frame	Q	Home
Jump to last frame	E	End
Next sequence	S	↓
Previous sequence	W	↑

Labeling:

Warning

Alternative keybinds for playback control do not work while an object is selected. They are used to control the ID field.

Action	Keybind
Select next object	F
Select previous object	Ctrl + F
Deselect object	R
Set ID in current frame	Enter
Set ID in all frames	Ctrl + Enter
Set ID in current and all following frames	Alt + Enter
Set ID in current and all previous frames	Ctrl + Alt + Enter

Utilities:

Action	Keybind
Backup ground truth files	Ctrl + B
Reorder IDs in sequence	Ctrl + R

Note

To be able to activate the Enter keybinds the focus must be in the ID field.

Credits

The original dataset RADDet which RADTrack is based on, was created by Ao Zhang, Farzan Erlik Nowruzi and Robert Laganiere from University of Ottawa and Sensorcortek Inc.

Citation

Please use the following citation when using the dataset:

@inproceedings{RadarMOTR2024,
    author = {Dell, Martin and Bradfisch, Wolfgang and Schober, Steffen and Klöck, Clemens},
    title = {{RadarMOTR: Multi-Object Tracking with Transformers on Range-Doppler Maps}},
    booktitle = {2024 International Radar Conference (RADAR)},
    year = {2024},
    pages = {1-6},
    doi = {10.1109/RADAR58436.2024.10994166}
}