Zeroth Law Traffic

Submission for the WIUT Hackathon 2026 computer vision track, built as team lead in a team of three. One fixed road camera, a few unlabelled sample videos, and a hidden annotated test set from the same angle.

Part A, event detection: given an .mp4, return every traffic event as [start_sec, end_sec, label] across 14 official classes: accidents, near misses, red-light running, wrong-way driving, illegal U-turns, stopped vehicles, jaywalking, failure to yield, illegal turns, solid-line crossings, stop-line violations, congestion, road obstacles, fire and smoke. Segments are matched by temporal IoU, so boundaries have to be tight.

Part B, accident anticipation: at every frame, the probability that an accident starts within the next 5 seconds, using only frames seen so far. The estimator never opens the video; the harness feeds it frame by frame, which keeps it causal.

Approach. CPU decode of every third frame, COCO-pretrained YOLO11m (no team fine-tuning), ByteTrack with Kalman filtering, rider fusion. A world model combines hand-mapped camera geometry, signal lamp colours, per-track kinematics and a flow atlas fitted to the samples, then one state machine per event class and segment post-processing. Part B runs its own detector and tracker, builds conflict features (closest approach, closing speed, braking, red-light and wrong-way modifiers) and turns them into an analytic risk score, smoothed and calibrated.

Part A and Part B share one budget: together they must finish within 3x the video's duration on a single T4, or both outputs for that video are discarded.

Honest status: the event engines stay disabled until they are validated against labels, so Part A currently emits no events. Risk calibration uses synthetic scenarios, not real accidents. T4 runtime is unmeasured; the historical CPU-laptop runs were 6-9x real time, not a 3x pass.

Stack: Python 3.11, Ultralytics YOLO11m, ByteTrack, PyTorch, OpenCV, imageio-ffmpeg