Development of a human following system using 3D LiDAR and ReID3D

Human-following is an important function for service robots, as it supports tasks such as mobility assistance and object transport. However, many existing systems do not handle cases in which the target person is hidden by obstacles, known as occlusion, and they often fail to continue tracking once the target is lost. To address this problem, we propose a human-following system that combines 3D LiDAR with a person re-identification method to enable re-detection and recovery after occlusion. For person detection, we use PointPillars, which provides fast and accurate inference, and we train it with the TAO Toolkit using a custom dataset of 19,500 point-cloud frames. For re-identification after tracking loss, we apply the ReID3D framework, which can extract strong features from point clouds and allows robust identification of the target. To evaluate the tracking performance of the proposed method, we conducted experiments using a public dataset. Furthermore, to assess the system's following performance in a real-world environment, we conducted experiments using a real robot with custom-designed routes and scenarios.