End-to-End Trainable Deep Neural Network for Radar Interference Detection and Mitigation

Preparation

1: Download the RaDICaL dataset and install the RaDICaL SDK https://github.com/moodoki/radical_sdk

2: Extract .h5 bags to numpy files using radical/extract.py /path/to/radical/radar_30m/*.bag.h5 /path/radical/radarcfg/outdoor_human_rcs_30m.cfg /path/to/output/radar_frames.npy.

3: Generate a train, val, and test split using python generate_training_data.py /path/to/output/radar_frames.npy /path/to/generated/files.

Training

First set required envorinment variables.

export NORM_PATH=/path/to/norm/pkl
export TRAIN_CLEAN=/path/to/generated/files/train_clean.npy
export VAL_CLEAN=/path/to/generated/files/val_clean.npy
export VAL_MASK=/path/to/generated/files/val_mask.npy
export VAL_DISTURBED=/path/to/generated/files/val_disturbed.npy

Depending on the method you want to train you can use different training scripts:

RIDAM Detection and Mitigation

• python train_ridam_detection_mitigation.py $NORM_PATH $TRAIN_CLEAN $VAL_CLEAN $VAL_MASK $VAL_DISTURBED

AE-Gate Detection

• python train_ae-gate_mitigation.py $NORM_PATH $TRAIN_CLEAN $VAL_CLEAN $VAL_MASK $VAL_DISTURBED

AE-Gate Mitigation

• python train_ae-gate_detection.py $NORM_PATH $TRAIN_CLEAN $VAL_CLEAN $VAL_MASK $VAL_DISTURBED

CNNTD Mitigation

• python train_cnntd_mitigation.py $NORM_PATH $TRAIN_CLEAN $VAL_CLEAN $VAL_MASK $VAL_DISTURBED

CNNRD Mitigation

• python train_cnnrd_mitigation.py $NORM_PATH $TRAIN_CLEAN $VAL_CLEAN $VAL_MASK $VAL_DISTURBED

Evaluation

Generate tests sets using generate_training_data.py. Use val_clean.npy as clean input and change simulation parameters according to your desired testing environment.

Detection Evaluation & Mitigation

To evaluate the detection and mitigation capability of different methods use the evaluate_detection.py and evaluate_mitigation.py, respectively. Inside you can comment/uncomment methods you want to evaluate. Please make sure that you set the ckpt path for the learned methods (commented by todo).

Checkpoints

will be released soon.