Project story
From question to working system.
This project applied cooperative perception methods in a real-world freight and port environment, where heavy trucks, forklifts, passenger vehicles, pedestrians, occlusion, and calibration create demanding conditions for reliable sensing.
A Mobile Unit for Sensing Traffic and NVIDIA edge-computing devices supported a modular, low-latency pipeline combining YOLOv8 detection, ByteTrack tracking, camera calibration, trajectory prediction, conflict estimation, and a live interface for safety managers.
Testing across right-turn, left-turn, and two straight-crossing scenarios produced average detection confidence scores of approximately 0.72 to 0.78. The evaluation also exposed an important limitation—ID switching during occlusion and complex turns—and documented geofencing and region-of-interest strategies to improve tracking continuity.