Clutter Detection in Automotive Radar Point Clouds Based on Deep Learning with Self-attention

As radar can directly provide the velocity of the targets in autonomous driving and is known for the robustness against adverse weather conditions, it plays an important role in contrast to camera and lidar. However, on the downside, radar is susceptible to ghosts or clutters, caused by several factors, e.g., multi-path propagation. The clutters can lead to erroneous object detection and cause severe traffic accidents in autonomous driving. Therefore, it is desirable to identify and remove anomalous targets as early as possible in application. In this paper, we present a novel network architecture based on PointNet++ to realize the clutter detection. The network aggregates three feature branches and applies self-attention to distinguish clutters from other detections. To sufficiently utilize the radial velocity and RCS, we cluster the point cloud by DBSCAN first and then extract local features of each cluster, such as mean value and RBF. Our method is evaluated on a real-world dataset, RadarScenes, and shows promising results for clutter detection.