Optimizing PointNet++ and DBSCAN for Object Detection in Automotive Radar Point Clouds

Enhancing object detection in automotive radar point clouds is vital for the advancement of Advanced Driver-Assistance Systems (ADAS). This paper introduces an innovative optimization of PointNet++ for semantic segmentation in conjunction with class-specific clustering using the DBSCAN algorithm, thereby addressing the inherent challenges of manual tuning and domain-specific design in traditional methods. Our approach leverages measured radial velocity for more effective sampling within PointNet++, augmented by comprehensive Hyperparameter Optimization (HPO) and Neural Architecture Search (NAS). This optimized PointNet++ demonstrates state-of-the-art performance in semantic segmentation on the RadarScenes dataset, achieving notable reductions in model size. Furthermore, we significantly enhanced object detection performance by introducing a scaling vector and applying HPO to fine-tune the DBSCAN algorithm's parameters for each object class.