Every traffic event, as a time segment — and an alarm before the crash.

What is learned, what is rule-based

learned Road-user detection only: YOLO11s pre-trained on COCO (Ultralytics weights, no fine-tuning — we have no labels for this camera).

rule-based Everything after detection: tracking, the scene map, signal reading, all 14 event classes and the Part B risk score. Rules are deterministic, explainable and unit-tested on synthetic trajectories drawn on the real map.

Why: the hidden test set contains events that are not in the samples, so any classifier trained on the samples would not have seen them. Rules written from the class definitions do not need examples.

Key design decisions

• Speeds in body lengths per second (image speed / √(box area)): removes most of the perspective, so one threshold works near and far.
• Scene registration. The framing moves between clips (up to 60 px and ~1° in our stills), so the map is re-registered on every video with SIFT + RANSAC.
• Only trust what is visible. The queue approaching the camera has its signal facing away, and the head we can see on that side is a pedestrian signal (it showed WALK while that queue was flowing), so red-light is judged only where the vehicle head faces the camera.
• Stopped vs queued: a stopped vehicle is one that other traffic keeps passing.

What we built

A detector → tracker → map → rules pipeline (Part A), a causal closest-point-of-approach risk model (Part B), and the tooling around it: scene registration, a labeling page for dev sets, annotated renders, EDA, a Gradio demo that runs on a free Colab GPU and 26 unit tests with synthetic trajectories on the real layout.

What worked

• SIFT registration fits the layout on every still, including dusk frames shifted by 60 px.
• COCO YOLO11s finds cars, buses, pedestrians and cyclists reliably in day and dusk frames.
• Calibrating on a labelled sample clip raised Score A from 0.265 to 0.705 (one clip, one annotator).
• Part B false alarms dropped from ~5 to ~1 per minute on the sample videos once overlapping (occluding) boxes and cars passing a queue in the next lane stopped counting as collision courses.
• All four sample videos run in about half the 3× time budget on a laptop CPU.
• Body-length-normalised speed made one set of thresholds work across the depth of the scene.

What did not (or is not proven)

• The thresholds come from one labelled 67 s clip and a detection-by-detection review of a second video — a first calibration, not a validated result.
• Reading the pedestrian head as the queue's signal was wrong (it showed WALK while that queue flowed); red-light for the queue towards the camera is therefore not judged.
• Top-down footage breaks COCO detection (cars read as "cell phones").
• Pixel cues (obstacle, fire) needed camera-drift compensation and shape filters to stop firing on lane markings; smoke without flames is not detected.
• illegal_turn needs the junction's prohibited movements; until confirmed it is never predicted, because a predicted class that never occurs costs a whole class in the macro-F1.
• Near-miss vs normal hard braking in a queue is the least certain boundary.

Next

• Label all four sample videos with the labeling page and tune thresholds per class at IoU 0.7.
• Fine-tune the detector on frames of this camera (dusk, occlusion in the queue).
• Learned re-scoring of accident / near-miss candidates on public crash datasets (CCD, DoTA).
• Calibrate the risk score on those datasets so 0.5 means "probably within 5 s".