Self-Supervised Class-Agnostic Motion Prediction for Autonomous Driving

Understanding the motion behavior of the dynamic environment is crucial for autonomous driving. Recently, there has been growing interest in class-agnostic motion prediction directly from point clouds. Since manual labeling of point cloud data is laborious and time-consuming, some self-supervised motion prediction methods have emerged, but most of these methods rely on multi-modal data. The inherent potential features of point cloud sequences deserve to be more fully explored. To this end, we propose to perform motion prediction based on pillar features of non-ground points and exploit the motion consistency among multiple frames of consecutive point clouds to facilitate the self-supervised learning. Experimental results on nuScenes demonstrate that our approach performs well in motion prediction of dynamic points.