How it works, and what we don't know yet

Given an MP4 from a fixed road camera, produce a list of traffic events as time segments — a start second, an end second and one of 14 class labels — and, separately, a per-frame score between 0 and 1 for the probability that an accident begins within the next 5 seconds.

The two outputs are scored independently, and they pull in different directions. Event detection rewards precise boundaries on things that have already happened. Anticipation rewards a causal signal that rises before anything has happened at all. Treating them as one problem produces a system that is late on both.

The organiser sample clips from a single fixed camera, plus the annotations released with them. Because every clip comes from the same viewpoint, scene geometry can be calibrated once and treated as known for all of them.

Class balance is the dominant property of this data: ordinary violations are common, collisions are rare, and fire or smoke may not appear at all. Any metric that averages over classes has to be read with that in mind.

Trajectories collapse onto two carriageway corridors plus a pedestrian crossing; motion density concentrates in a small fraction of the frame. Both findings argue for spending effort on scene calibration rather than on generic whole-frame modelling.

Lighting varies across the sample set from midday to night. The night clip carries noticeably fewer tracks at a comparable apparent density, which is the signature of a detector dropping small, poorly-lit road users rather than of a quieter road.

The full analysis, with the decision each chart led to, is on the EDA page of this site.

Decode at a fixed stride, detect road users, track them into trajectories. After tracking the pipeline forks: one branch applies scene geometry and traffic rules to produce event segments, the other computes interaction features between track pairs to produce the risk curve.

The fork matters for debugging. A missed event and an under-confident risk curve have different causes and are fixed in different places, and keeping the branches separate keeps that distinction visible.

A single-stage detector proposes boxes for six classes: car, truck, bus, motorcycle, bicycle and pedestrian. Input resolution is chosen by the smallest pedestrian that has to survive, which in practice is set by the night clip rather than by any daytime footage.

Detections are associated across frames so each road user keeps one identity for as long as it is visible. Short occlusions are bridged rather than allowed to split a track, because a split track breaks every rule that reads trajectory history.

Identity stability, not raw detection accuracy, is the property most rules actually depend on.

The road polygon, lane polygons with expected directions, stop lines, crosswalk, intersection box, solid markings and the signal region are calibrated once for the camera and reused across clips.

This is the cheapest accuracy available in the whole system. It converts several classes from open vision problems into geometry tests, and it makes failures inspectable: a wrong-way false positive usually means a lane vector is wrong, not that a model is confused.

Where the traffic code has a crisp definition, the definition is encoded rather than learned: wrong-way driving as a sustained heading opposite the lane vector, stopped vehicle as displacement under a threshold for at least ten seconds inside the road polygon, jaywalking as a pedestrian centroid inside the road and outside every crossing, solid-line crossing as a trajectory intersecting a calibrated line.

Rules are deterministic, cheap and inspectable, and each one names the geometry it depends on. That is what makes a false positive traceable to a specific polygon.

Collisions and near misses resist a clean geometric definition, so they are decided from interaction features over track pairs: relative motion, path intersection, time-to-collision, and the deceleration and heading changes around the moment of conflict.

The distinction between the two classes is what happens after the conflict. A near miss is a low time-to-collision where both tracks survive and never overlap; a collision is where they do.

The risk head answers one question per frame: what is the probability that an accident starts within the next 5 seconds. It is causal by construction — it may only read frames up to the current one.

Time-to-collision is the strongest available signal, but it over-fires in dense traffic where everything is already slowing. Deceleration, heading change, lane context and local density are combined with it so that ordinary congestion does not read as imminent collision.

The property to check is lead time, not peak height: a curve that spikes after impact has detected a crash rather than anticipated one, and would be worthless in operation even if it scored well.

Per-frame decisions are noisy, and the evaluation reads segments. Fragments of one class are merged across short gaps, segments below a minimum duration are dropped, boundaries are snapped to where evidence actually begins and ends, and two segments of the same class are never allowed to overlap.

This stage changes scores more than most model choices do, because it is what converts a correct-but-flickering signal into a segment that matches an annotation.

The ablation table below is the intended experimental frame: measure the contribution of tracking, scene rules and temporal smoothing separately, so the score can be attributed rather than guessed at.

Inference runs offline on the evaluation machine with no hosted AI service. Frame stride is the main runtime control: detection dominates cost, and the stride sets how often it runs.

No real-time claim is made anywhere on this site, because no runtime measurement has been taken yet.

The categories below are the ones this design is structurally vulnerable to, listed before measurement rather than after, so they can be checked rather than rationalised.

Calibrating the scene once and treating geometry as known — it removed whole classes of ambiguity and made rule failures traceable.

Keeping detection of events and anticipation of accidents as separate branches over shared tracking, instead of one model asked to do both.

Building the visualisation before tuning thresholds. Seeing a risk curve next to its clip makes a bad threshold obvious in seconds, where a metric only says the number moved.

Multi-camera calibration, so the same geometry approach extends to a network rather than one view.

Explicit signal-state estimation rather than reading colour from a fixed region, which is the most fragile input in the current design.

An operator feedback loop: let a human confirm or dismiss an alarm, and use those decisions to retune thresholds per camera.

Any pretrained weights, external datasets and third-party libraries used are listed here with their licences, so the submission can be reproduced and audited.

The intended path is: clone the repository, install pinned dependencies, download weights, run inference on a clip, and get the same predictions file that produced the results on this site.