Deep Learning Based Target Detection Method for the Range-Azimuth-Doppler Cube of Automotive Radar

Automotive radar has an efficient ability to estimate velocity and adapt to adverse environments. Range-Azimuth-Doppler (RAD) cube as a representation of radar data, adequately retains the original information and has been used for detection tasks based on deep learning. However, in the prior work, the overall target feature in the RAD cube is not fully utilized, as the feature of each dimension is extracted separately or unevenly. In this paper, we propose a triple-head detection network (TriHeadDet) using RAD cube as input. Firstly, the overall feature is extracted by 3D convolution. And then a detection task is implemented in range, azimuth and Doppler dimension, called the 3D detection in this paper. Specifically, the 3D detection task is decomposed into 2D ones with shared neck feature and parallel detection heads. A convenient bounding box fusion method is proposed to achieve the final 3D detection. Through experiments on simulated and measured datasets, the proposed method shows the best performance among all experimental methods, with relatively high detection probability and low false alarm probability. In addition, the proposed network has a certain generalization ability for low signal-clutter ratio (SCR) and multi-target scenarios.