Efficient Overhead People Counting via Improved Track Association and GIoU-Based Similarity for Public Transport Monitoring

In public transportation, vision-based monitoring counting systems can contribute to improved safety and resource management within public transportation systems. In this study, we introduce a novel Line-of-Interest (LOI) counting model tailored for people counting from an overhead view in video footage. Specifically, an improved similarity matrix is used during the association phase to align detections with tracks. In addition, the integration of the Generalized Intersection over Union (GIOU) matrix with distance information of detection establishes a new similarity matrix, which effectively decreases the effects of occlusions and variations in bounding boxes in counting problems from the overhead perspective. The experiments achieve better results compared to the existing LOI(Line of Interest) methods in evaluating publicly available datasets. This advancement represents a substantial improvement in precision and efficiency, demonstrating the enhanced capability and robust performance of the proposed model in comparison to established benchmarks. Furthermore, the proposed approach is comprehensively evaluated through the use of recall, precision, and F1 score metrics. The experimental conclusions indicate that the proposed model notably enhances the efficiency of the counting process without causing additional computational expenses. This renders it well-suited for deployment in embedded systems to enable real-time people count in public transportation settings.