Anti-low angle resolution: Some adaptive improvements in anchor-based object detection algorithm for automotive radar target detection

Automotive radar has extensive and well-established applications in advanced driver assistance systems (ADAS). It has been the most reliable and preferred sensor for functions such as adaptive cruise control and automatic emergency braking. Nevertheless, the limited angle resolution of traditional millimeter-wave radar data reduces its effectiveness in autonomous driving systems (ADS). Our study presents some adaptive improvements in anchor-based object detection algorithm framework to address the inaccuracies in target detection and positioning resulting from the limited angle resolution of millimeter-wave radar data. Our model utilizes the Range-Angle (RA) and Range-Doppler (RD) views of a RAD cube as inputs. An anchor assignment strategy- Max Intersection over Union (IoU) assigner is used to generate more positive samples than before, enabling the model to learn more about the relevant features of targets. We propose a multi-detection box fusion Non-Maximum Suppression (NMS) mechanism to fuse prediction results with varying confidence scores, thus improving the accuracy of target detection positions by addressing the issue of target box position deviation caused by low angle resolution. In addition, an actual distance loss function is designed to constrain the regularity that the size of the bounding box is inversely proportional to the distance between the target and the radar. Based on experimental results, the proposed method increases the mean average precision (mAP) by approximately 10% compared to the baseline method.