Learning Semantics on Radar Point-Clouds

Localization and perception research for Autonomous Driving is mainly focused on camera and LiDAR data, rarely on radar data. We apply an automated labeling pipeline to semantically annotate real world radar measurements, manually correct point-wise labels to obtain ground-truth, and apply supervised learning models on this data. To assign an attribute, called class label, to every point of an input cloud is hereby referred to as semantic segmentation. Transferring approaches of LiDAR segmentation into the similar data structure, we research deep-learning semantic segmentation on radar point clouds. Compared to classical Cartesian coordinates, a polar coordinate input discretization benefits the dynamically changing number of radar detections per sensing cycle and simplifies to model the quasi-radial sensor resolution. Moreover, we evaluate different network architectures, examine radar feature channels and also temporal consistency by attention map concatenation. Our contribution is twofold. First, featuring a semantically labeled real world radar dataset for ground truth. Second, our supervised learning approach to solve semantic segmentation on radar point-cloud data. Our classification benchmark network yields 56.1 % weighted Intersection of a Union of relevant classes for radar, while reaching a real-time framerate of 12.4ms.