A Change Detection Method for Misaligned Point Clouds in Mobile Robot System

3D change detection of point clouds measured by 3D sensors such as LiDAR is become more widespread and is now being used in various tasks related to natural disaster evaluation, agricultural assessment, and so on. Such tasks are increasingly being automated by mobile robots such as Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs). For efficient task execution with robots, real-time change detection is required. Some degree of localization error of robots when performing tasks is inevitable, and this makes it difficult to align point clouds measured at different times. The point clouds need to be converted into a global coordinate system in order to compare them, and the misalignments caused by localization error degrade the accuracy of the change detection. In this paper, we propose a change detection method that is robust to the misalignment of point clouds. Our method does not require training and takes a feature-based approach utilizing occupied, free, and unknown attributes created by analyzing the point clouds. Using a realistic robotics simulator, we generate synthetic point clouds of warehouse scenes for change detection tasks, and the misalignment of point clouds is reproduced. We evaluated the proposed and conventional methods with the point clouds, and the proposed method proved to be more accurate than conventional methods.