GateTrace: Where Do Crowd Crossings Go Missing Under Occlusion?

📅 👥 3 authors arXiv:2611.26388v1

Authors

Abstract

At large events, pedestrian tracking supports entrance and passageway counts, but totals hide which crossings were missed. In a stadium study, the system recorded about four in five clearly visible crossings but fewer than one in three mostly obscured crossings. The videos lacked trajectories annotated across frames, so we turned to MOT20, which provides pedestrian trajectories and visibility. We introduce GateTrace to locate missed crossings and assess changes to counting rules. It derives reference crossings from annotations under fixed rules, matches reports by person, direction, time and position, and traces misses through saved boxes and counting decisions. With a shared detector, ByteTrack, OC-SORT and BoT-SORT recover 38.9-46.7% of low-visibility reference crossings on development sequences, versus 98.1-99.0% at high visibility. The gap recurs on two further sequences under a protocol fixed in advance. Published outputs of MOT20-tuned BASE recover 63-78% at low visibility across both sequence groups. Losses stop at different stages across systems. Across four systems and both groups, 47-79% of low-visibility misses have no nearby output linked to the person or no possible crossing in the saved tracks. Removing later checks from BASE on development sequences leaves most misses unresolved and lowers precision against the fixed reference list from about 96% to 77%. Misses also remain among sampled reference crossings that two human annotators supported after checking the videos. GateTrace shows where crossings are lost and measures the recovery and precision costs of changing later checks. We will release the code and per-crossing records.

Paper Information

arXiv ID:
2611.26388v1
Published:

Related Categories