Real-Time Transit O-D Data Collection via Edge-AI-based Passenger Re-Identification

Abstract: This paper presents a novel approach for collecting origin-destination (OD) data in public transit systems using edge computing and computer vision. Traditional methods for gathering OD data, such as manual surveys, Bluetooth, or automatic fare collection systems, often prove inefficient, labor-intensive, and raise privacy concerns. To address these challenges, we propose a framework that leverages existing onboard cameras in public transit vehicles to estimate OD data by tracking and re-identifying passengers in real-time. The system operates on edge devices, processing video locally without transmitting raw footage to the cloud, thus minimizing data transmission and protecting passenger privacy. By using edge computing, data generated by field sensors, such as video feeds, are processed in real-time, improving efficiency and reducing bandwidth usage. The framework employs YOLOv9 based detection and tracking followed by Global-guided Reciprocal Learning (GRL) for accurate passenger re-identification, classifying their trajectories as either boarding or alighting, which helps build the OD matrix for the transit system. We implemented and tested this framework on the NVIDIA Jetson AGX Orin, an edge device suitable for real-time processing. The framework was evaluated using real-world bus surveillance videos captured over a full day. Despite challenges such as varying lighting conditions, occlusions, and the lack of domain-specific datasets, the proposed system achieved a rank-10 accuracy of 76.3% in passenger re-identification. This result is promising for applications in real-world transit environments, where consistent and high-quality OD data are crucial for optimizing transit operations. Future research will involve deploying the system onboard buses to test its real-time performance in live conditions. Additionally, gathering more diverse and domain-specific datasets is expected to further improve the accuracy of passenger re-identification, particularly in complex scenarios like chained trips where passengers switch between buses. Challenges remain in recognizing passengers using mobility aids like wheelchairs, which may cause occlusions during boarding and alighting. Addressing these issues will require the development of training datasets representative of diverse passenger demographics, as well as different lighting and viewing conditions. In summary, this study demonstrates the potential of edge AI-based computer vision systems for real-time OD data collection in public transit. The framework effectively mitigates privacy concerns while providing transit agencies with accurate and efficient tools for route planning, scheduling, and service optimization.