Enhancing Millimeter-Wave Radar Object Detection with a 3D Transformer Module on Range-Azimuth Tensors

The millimeter-wave (mmWave) radar has been a robust and cost-effective sensor in the perception system. However, mmWave radar object detection is still a great challenge because radar signals contain far less semantic information than RGB images and are difficult to decipher directly by a human observer. Considering this, this letter presents an end-to-end network termed as 3D-Transformer-CNN (3D-T-CNN) for mmWave radar object detection using Range-Azimuth (RA) tensors. In this method, a non-local module transformer based on multi-head self-attention is introduced into a 3D vanilla convolutional neural network (CNN), which can make the best use of local and global features of sequential RA tensor frames. The experimental results show that the 3D-T-CNN significantly improves the accuracy of radar object detection, which achieves state-of-the-art performance.