Detection and Tracking in Human Monitoring Framework Using Modified Direct 3D LiDAR Point Cloud Classifier Based on Region Cluster Proposal

Using camera pixel data in visual data monitoring raises privacy issues as it captures the entire environment and sensitive information. Hence, numerous studies have investigated human monitoring procedures using Three-dimensional Light Detection and Ranging (3D LiDAR), specifically focusing on detection and tracking tasks to mitigate potential health risks from positional patterns. Unfortunately, unprocessed 3D LiDAR point clouds are challenging to detect and track due to their dispersed nature. As suggested in many studies, restructuring strategies effectively decrease information loss but require higher computational costs. This study proposes the utilization of a direct point processing approach based on a region cluster proposal on a modified PointNet classifier as human detection and tracking framework. The modified PointNet classifier has demonstrated an improved accuracy of 98.79% in the classification process for supporting human object detection, higher than the default architecture which yields an accuracy of 94.98%. Furthermore, this study also develops a distance estimation technique to enhance the tracking process. In general, the human detection and tracking procedure demonstrates satisfactory performance through the utilization of a solitary data type, specifically the unprocessed 3D LiDAR point cloud, which is processed directly using a modified PointNet classifier.