RMSA-Net: A 4D Radar Based Multi-Scale Attention Network for 3D Object Detection

Abstract

Perception is a fundamental task in autonomous driving systems(ADS) and robotic systems alike, which commonly uses multi-modal sensors including cameras, LiDAR, and radar. Recently, 4D radar has gained increasing attention thanks to its advantage of high angular resolution in both azimuth and elevation. As the unique features of 4D radar point clouds call for more efforts on brand new object detection algorithms based on 4D radar, there are few studies in this topic. Therefore, in this paper we fully consider the characteristics of the sparse and shapeless 4D radar point clouds and propose a model RMSA-Net to fuse rich contextual information and address the problem of height information loss. We conduct extensive experiments to demonstrate that our model achieves the gain in terms of 3D mAP and bird’s eye view(BEV) mAP by 3.28% and 0.96%, respectively.