An Anchor-Free Dual-Branch Approach for Real-Time Metro Passenger Detection

Accurate and real-time detection of metro passengers is of utmost importance in ensuring public safety and facilitating efficient transportation management. However, this task remains challenging due to spatial constraints, complex backgrounds, scale variations, and mutual occlusion. In this article, we propose AMPDet-an anchor-free metro passenger detection model that features a dual-branch architecture. This architecture is designed to concurrently identify both the head and body regions of passengers. By leveraging this dual-branch architecture, AMPDet effectively prioritizes genuine pedestrian characteristics during its decision-making process, thus mitigating the intrusion of nonpedestrian elements. Additionally, we propose a head-body joint nonmaximum suppression (NMS) technique that combines the outputs of the two branches, aiming to efficiently recover suppressed body detections. Comprehensive experiments on CityPersons, CrowdHuman, and SH-metro show the superior performance of the proposed method in pedestrian detection, especially in crowded metro carriage scenes. Further, AMPDet showcases exceptional adaptability and efficiency on the NVIDIA Jetson AGX Xavier platform, affirming its practicality for implementation in metro passenger detection. Code and dataset will be available at https://github.com/weimingqi111/ SH-metro.