Real-time Moving Object Segmentation with Tracking and Tracklet Belief

Moving object segmentation (MOS) is a challenging but essential part for autonomous robotic systems. Segmentation of moving objects provides significant information for mapping, avoiding collisions, and planning task. In this work, we address the problem of object detection and tracking in 3D LiDAR sequences. Our approach achieves robust moving object segmentation and object velocity estimation by unique object clustering and association strategy. An object-level dynamic belief is calculated to remove incorrect associations and non-moving objects. Experiment results on SemanticKITTI dataset show that our approach accurately detects moving objects and outperforms state-of-the-art MOS methods. Furthermore, we verify our approach on a self-collected campus dataset with Velodyne VLP-16. The results show that it adapts well to different sensor setups since it incorporates information of objects’ motion and does not depend on training data.