Technical Approach: Vision-Based Passenger Counting Prototype
Hichem Lakhdhar, Muhammad Zohaib · Zenodo (CERN European Organization for Nuclear Research) · 2026
Accurate knowledge of passenger numbers aboard a vehicle and their boarding times are essential for transit operations. Operators depend on this data to plan schedules, determine fleet size, forecast demand, and distribute revenue across routes. Historically, passenger counting relied on manual methods—staff using hand counters or periodic surveys. Automated Passenger Counting (APC) systems replace these manual approaches with continuous sensor-based monitoring. Multiple sensor technologies are available for this purpose, including infrared gates, pressure sensors, weight-based systems, and cameras. Among these options, camera-based systems capture the most detailed information. However, they also demand the most sophisticated analysis, as the boarding area during passenger entry presents a complex, crowded scene that easily confuses detection systems. The central question driving this study is straightforward: Can we build a detection system that is fast and efficient, combined with a transformer model used carefully and selectively, to count passengers reliably while processing video in real time? Three contributions came out of it. The first is a labelling and training routine that never asks a person to draw a box. The second is a design that splits the problem in two on purpose: the network learns the part that survives a change of viewpoint, namely whether a figure is moving, while the viewpoint-bound part, which way it is going, is handed to plain geometry. The third is an evaluation that does not flatter the result, reporting what the system achieves and placing it honestly besides what the APC literature reports.