Accuracy & error risk — honestly

Bottom line: this is a screening tool. From one uncalibrated camera you cannot state a defensible "violation rate". So: counts are estimates, events are candidates, and true accuracy is computed from human verification.

Where the AI is wrong, and why

• Occlusion: at a crossing trajectories intersect — a pedestrian hidden by a vehicle is the hardest case; the tracker loses the ID, the count undershoots.
• Single-camera perspective: foreshortening makes a vehicle still short of the stop line appear to overlap the crossing in image space — the biggest source of false positives.
• Small objects / panorama: a distant pedestrian is tens of pixels — small-object accuracy is routinely 2–3× lower (COCO).
• Night, rain, glare: out-of-distribution imagery — a steep accuracy cliff; we treat those periods as low-confidence.
• No metric depth: without homography we work in pixels — hence uncertain speeds and "did it stop in time".

What to realistically expect

• Daytime pedestrian/vehicle detection on a good view: useful but imperfect (small objects and crowding lower recall).
• "Failure to yield" from a single uncalibrated camera: many false positives — which is why human verification is mandatory.

How we compute true accuracy

Every flagged event has a snapshot. Humans vote confirm/refute. Precision = confirmed / judged. That figure is shown live on the home page and grows trustworthy as votes accumulate.

What pushes it to the maximum

1. Metric calibration (homography) — biggest win: unlocks speed, PET/TTC, correct "did not yield".
2. Higher fps + temporal smoothing — fewer ID switches.
3. Bigger model / camera choice (tight framing) — higher recall.
4. Active learning from human votes — the model learns from its own mistakes.
5. Restricting claims to what's reliable — honesty as an advantage in front of an audit.