Panoptic Segmentation for Automotive Radar Point Cloud

This paper presents an efficient approach to point cloud panoptic segmentation for automotive radar. Unlike most existing point cloud panoptic segmentation works, this method directly predicts both instance ID and semantic class without such intermediate step as clustering and is end-to-end differentiable. This is achieved by a novel instance labeling scheme in the training stage that facilitates the direct prediction of instance ID in the inference stage. More importantly, this new approach has low computation load and can achieve high frame rate in implementation, which is crucial for real time applications in automotive radar. Based on experimental evaluation with existing lidar point cloud data and our own radar point cloud data, we demonstrate that the proposed panoptic segmentation method can achieve good performance at a frame rate of higher than 150 FPS.