Point Cloud Instance Segmentation and Classification for Roadside LiDAR

Point cloud segmentation is an essential and efficient task in high-definition trajectory-level traffic information extraction for roadside LiDAR. State-of-the-art methods use conventional step-by-step in-sequence or all-in-one deep learning methods to process point cloud data. While effective in many applications, their deployment in complex traffic scenarios often faces constraints of computational resources and real-time requirements. To complement these approaches, a novel point cloud segmentation method was proposed jointing with background filtering and object classification to segment the objects in a single point cloud frame. First, a novel multi-scale mapping clustering algorithm was introduced for instance segmentation to address the issue of over- and under-segmentation caused by occlusions in traffic scenarios. A passage space feature was then proposed to better characterize the properties of instances in traffic scenarios. Finally, a local hierarchical classifier was employed to enhance the accuracy of object classification. Experimental results showed that the proposed point cloud segmentation method achieved more than 80% homogeneity, with the object classification accuracy increasing to 98%. Furthermore, the data processing rate is 8.67 ms per point cloud frame. These experimental outcomes highlight the robustness, accuracy, and efficiency of the proposed method in point cloud segmentation.