Case study - AI passenger-flow analytics for city buses

Passenger counting is not a single-frame detection problem. The system must understand door areas, trajectories, time windows, and state changes while coping with occlusion, vibration, lighting changes, and crowded scenes.

The result also needed to integrate with a real operational workflow through task submission, asynchronous processing, status tracking, callbacks, logs, and recoverable failures.

We designed a pipeline covering passenger and door detection, stable door-zone modeling, BoTSORT and ReID tracking, trajectory matching, boarding and alighting event recognition, and OD association.

Each key event keeps structured intermediate results and evidence frames so abnormal statistics can be traced back to the task, clip, trajectory, and decision rather than rerunning a black-box script.

The models were packaged behind APIs with asynchronous queues, workers, caching, task state management, callbacks, monitoring metrics, and deployment documentation.

We also adapted inference for Ascend 300i ACL, including model outputs, buffers, numerical edge cases, API compatibility, and performance profiling.

The final system can be integrated into a transportation business platform and operated as a maintainable production service.

The same architecture can extend to multi-stream RTSP input, occupancy analysis, station-level flow, abnormal behavior alerts, and industrial or facility video analytics.