Pedestrian Tracking Using 3D LiDARs - Case for Proximity Scenario.

In this paper, we address a critical issue of pedestrian tracking using 3D LiDARs. In particular, the segmentation of 3D point clouds to identify individual pedestrians is not stable when multiple pedestrians walk in close proximity (e.g., walking in a group), which significantly degrades tracking accuracy. To cope with this issue, we introduce a CNN-based method that predicts the population in each 3D point cloud segment. The Kalman-filter-based tracking process is fully tailored to include this prediction. We evaluated the proposed method in both a crowded environment and a real-world scenario using 3D point clouds obtained from multiple LiDAR units installed in our university campus building. The evaluation resulted in a MOTA (Multiple Object Tracking Accuracy) score of 0.977, representing the accuracy of person detection and ID assignment in tracking multiple individuals.