Dynamic Point Cloud Segmentation System for Under-Train Inspection Robot*

The under-train inspection robot operates inside the inspection pit under the metro train, precisely collecting point cloud and image data of critical under-train components to analyze the health of the train. The robot employs laser Simultaneous Localization and Mapping (SLAM) for autonomous localization. The inspection pit is a shared environment for both humans and robots. During operation, the environment is dynamic with possible worker collaboration. Personnel movement generates dynamic point clouds, which affect SLAM systems that assume a static environment. This reduces precision and affects operational safety and inspection effectiveness. To address this challenge, this paper introduces a dynamic point cloud segmentation system, suitable for the low-cost hybrid solid-state LiDAR used by the robot. The system utilizes appearance and motion features in 2D Bird's Eye View (BEV) to separate dynamic point clouds from input data. It also integrates spatial density clustering results to enhance segmentation and improve the accuracy of dynamic point recognition. Experiments in a simulated inspection pit environment demonstrate the system's effectiveness. The results indicate that the system effectively recognizes and segments dynamic point clouds. This capability allows for generating scan sequences that filter out dynamic human figures, thus minimizing SLAM interference from dynamic point clouds.