This is the official implementation of DAROD paper (2022 IEEE Intelligent Vehicle Symposium).
Colin Decourt, Rufin VanRullen, Thomas Oberlin, Dider Salle

Please cite this paper as follows:

@INPROCEEDINGS{9827281,
  author={Decourt, Colin and VanRullen, Rufin and Salle, Didier and Oberlin, Thomas},
  booktitle={2022 IEEE Intelligent Vehicles Symposium (IV)},
  title={DAROD: A Deep Automotive Radar Object Detector on Range-Doppler maps},
  year={2022},
  volume={},
  number={},
  pages={112-118},
  doi={10.1109/IV51971.2022.9827281}}

In this repository we provide an adaption of the Faster R-CNN architecture for object detection on range-Doppler (RD) maps.
We provide training and evaluation scripts for Carrada and RADDet datasets.

This project requires tensorflow 2.4.1 and tensorflow-datasets 4.4.0 with CUDA 11.2. Clone the project:

git clone ...
cd darod

Create conda environment
conda create -n darod python=3.7 tensorflow-gpu=2.4.1 scikit-learn scikit-image matplotlib numpy scipy tqdm tabulate imageio

Install missing packages
pip3 install tensorflow-datasets==4.4.0 tensorflow-addons==0.13.0

For efficiency, we use tensorflow datasets to load and process data. Before training the model, create tensorflow datasets for Carrada and RADDet.

Update path to Carrada dataset in datasets/carrada_builder/carrada.py (l 87):

def _split_generators(self, dl_manager: tfds.download.DownloadManager):
    """Returns SplitGenerators."""
    # TODO(carrada): Downloads the data and defines the splits
    path = "<path_to_carrada>"

    # TODO(carrada): Returns the Dict[split names, Iterator[Key, Example]]
    return {
        'train': self._generate_examples(path, 'train'),
        'test': self._generate_examples(path, 'test'),
        'val': self._generate_examples(path, 'val'),
    }

To create tf-records file for Carrada, use the following commands:

cd datasets/carrada_builder/
tfds build carrada --data_dir <path_to_tensorflow_datasets>
cd ..

Update path to RADDet dataset in datasets/raddet_builder/raddet.py (l 61):

def _split_generators(self, dl_manager: tfds.download.DownloadManager):
    """Returns SplitGenerators."""
    # TODO(carrada): Downloads the data and defines the splits
    train_path = "<path_to_raddet>/train/"
    test_path = "<path_to_raddet>/test/"
    # TODO(carrada): Returns the Dict[split names, Iterator[Key, Example]]
    return {
        'train': self._generate_examples(train_path, 'train'),
        'test': self._generate_examples(test_path, 'test'),
    }

Build the dataset:

cd raddet_builder/
tfds build raddet
cd ../../

Once the datasets are generated, update configuration files for Carrada and RADDet. Open config_carrada.json and config_raddet.json and change the following lines:

"dataset_version": "1.0.0", # 1.0.0 by defaults
"tfds_path": "<path_to_tfds_folder>", # <path_to_datasets>/tensorflow_datasets/

python train.py --config ./config/config_carrada.json --backup-dir ./logs/ --exp darod_carrada

python train.py --config ./config/config_raddet.json --backup-dir ./logs/ --exp darod_carrada

We provide DAROD pretrained weights. You can use them to evaluate the model.

We store the results in a .json file such that:

"class_id": { # metrics for class class_id 
  "AP": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
  "precision": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
  "recall": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
  "F1": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],

...

"mean": { # mean of metrics over all classes
  "AP": [
      [email protected],
      [email protected],
      [email protected],
      [email protected],
  ],
  "precision": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
  "recall": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
  "F1": [
    [email protected],
    [email protected],
    [email protected],
    [email protected],
  ],
}

Supposing, logs are in ./logs/darod_carrada:

python eval.py --path ./logs/darod_carrada

Supposing, logs are in ./logs/darod_raddet:

python eval.py --path ./logs/darod_raddet

This implementation repository is based on the following Faster R-CNN implementation: https://github.com/FurkanOM/tf-faster-rcnn. Thanks to the authors for the work.