Size-Prior-Oriented Target Detection and Recognition for Automotive SAR

Zekang Fan, Bo Zhao, Cuiqi Si, Fengming Huang, Qiuchen Liu, Lei Huang

Abstract: Automotive SAR target detection, which involves interpreting scenes or distinguishing different objects from SAR images, is a fundamental and critical problem in intelligent driving. An increasing number of methods have been proposed in airbone-SAR image understanding due to the challenges in deficient and high-variable SAR samples. In the context of automotive SAR, beyond these challenges, the specific incidence angle of radar scattering mechanisms in millimeter-wave band present additional difficulties in target identification. Therefore, this article proposes a prior-guided-attention module, termed as size oriented module, based on the backbone of YOLOv5. Then, with the newly established automotive SAR image dataset, amounts of experiments in the open world are conducted. the false and missed recognitions were reduced and the mean average precision (mAP) improvement of each method was about 3% . A test result mAP of 92.8% was achieved on the real-measured data, and the role of the individual modules was analyzed with the help of gradient-weighted class activation mapping and test results, thereby the effectiveness of SM with attention module is verified.