Results on the sample videos

Every sample, annotated by the submitted code: boxes with track IDs and speeds, the scene overlay, banners for active events and the risk bar. Click an event in the timeline or the table to jump to it.

One example of each class we detect

A still from our annotated video in the middle of the event. "Checked" means we verified this event frame by frame against the task's definition; click to play it.

Failure cases we know about

• Two people walking side by side on a crossing were flagged as a near miss (C3902, 13 s), and a person waiting at the far kerb for 7 s as jaywalking (C3902, 166 s). Both rules were fixed: a near miss needs a motor vehicle, a jaywalker must walk at least 1.5 body lengths on the road.
• Queues regularly spill past the stop line into the junction at noon. A car caught in such a queue can be reported as a stop-line violation while its own signal has already turned green (C3896, 60 s).
• Far-field pedestrians at the opposite kerb are small (under 3% of the frame height); the rules ignore them, so events on the far side of the road are missed.
• Dense, shallow-angle traffic: closer vehicles hide the bottom of farther ones, so box bottoms jump. We take velocity from the box top, which cut 1-in-1000 acceleration spikes from 4.1 to 1.6 body lengths/s² on the highway clip, but near-miss candidates in stop-and-go traffic still reach 0.63 confidence against a 0.70 gate.
• Far-field objects under about 3% of the frame height have noisy speed and heading. Risk features and near-miss pairs ignore them, so conflicts far from the camera are missed.
• Signal-dependent classes (red light, stop line) need the signal head inside the frame and a drawn stop line. Without them the rules switch themselves off rather than guess.
• ID switches after occlusion: tracks are split at implausible jumps, and stationary vehicles are re-linked by position, but a pedestrian hidden behind a bus becomes two tracks.