Doppler Beam Sharpening for 3D Object Detection

In this paper we investigate the use of Doppler Beam Sharpening (DBS) as a preprocessing step for radar data in an autonomous driving 3D detection pipeline, enhancing the effective angular resolution. Due to the assumption of a static scene, the proposed method applies DBS together with a moving point filtering approach to reduce preprocessing artifacts for dynamic objects. We derive an optimal correction region to apply DBS selectively at angles where it provides a resolution advantage over the base angular resolution. We show promising quantitative results, with an average improvement of 5.2% in average precision (AP) for off-boresight objects and run an ablation study isolating the effects of the individual components of our approach. While our experiments use a conventional Region Proposal Network (RPN) detection model with radar and camera Feature Pyramid Network (FPN) backbones, the approach is applicable to any radar object detection pipeline.