Non-parametric Dynamic Point Cloud Segmentation System Based on Time Series and Spatial Clustering

This paper tackles the challenge of robust LiDAR SLAM in dynamic environments and introduces a non-parametric dynamic point cloud segmentation system based on time series and spatial clustering. This system can identify and eliminate dynamic point clouds from the input point cloud frames without depending on any parameters or learning, thus obtaining static point cloud frames for building the SLAM system. This paper employ a non-parametric network to encode the point cloud into high-dimensional features and conduct similarity matching through a memory bank to assign classification labels to each point cloud point. We devise a result optimization method based on a sliding window and spatial clustering, which leverages the dynamic correlation of adjacent frames in the time series to holistically optimize the segmentation results, thereby more precisely detecting and filtering out dynamic point clouds. In the real industrial production environment, this paper performed experimental evaluations of the dynamic point cloud segmentation system proposed, verifying its efficacy and robustness in dynamic scene SLAM. Compared with other algorithms, this paper achieves a higher mIoU (0.881) and segmentation accuracy (0.957). The method of this paper can reduce the pose estimation error of the SLAM system by removing dynamic objects, enhance the quality of the map, and prevent positioning failure and other issues. Since no training is required, the method of this paper can adapt to different scenes and dynamic objects without tuning parameters or relearning, and has strong generality and flexibility.