Concepts

Zones

A zone is a polygon an operator draws once. From then on it belongs to the ground underneath it, not to the screen. The camera moves; the zone stays on its patch of world.

This distinction sounds small and is everything. A screen-space region silently changes meaning every time the drone drifts a meter. “Count cars entering this area” is only a coherent instruction if the area refers to a fixed place, so anchoring is treated as a core competency of the runtime rather than a rendering detail.

How anchoring works#

Capture, at creation

When the operator finishes drawing, the runtime samples the region: distinctive keypoints, a spread of trackable corner features, and a small template image of the patch itself. This bundle is the zone's visual fingerprint.

Follow, every frame

Optical flow tracks the feature points from frame to frame and re-derives the polygon from their motion. Cheap, sub-millisecond, and accurate while the scene moves smoothly.

Reacquire, on failure

Flow breaks under fast pans, motion blur, or occlusion. The runtime then falls back to matching the stored keypoints against the current frame, and past that, to searching for the template patch directly. Whichever succeeds re-seats the polygon.

Sanity-check, always

Every proposed update has to pass plausibility tests before it's accepted: the implied transform must be geometrically reasonable, and the patch under the new polygon must still resemble the fingerprint (similarity above 0.64). A zone that can't verify itself reports lost instead of guessing.

With telemetry available, zones can additionally be grounded through camera pose, which holds up under moves that pure vision can't follow. Both modes coexist; vision is the baseline, telemetry is reinforcement.

Failure honesty#

The design rule for zones is that a wrong answer is worse than no answer. A zone that has lost registration greys out and stops producing events until it reacquires. It does not slide to wherever the pixels went. Operators learn to trust the overlay precisely because it admits when it doesn't know.

See one working#

The challenge uses a single anchored zone on real boulevard footage with a moving camera. Watch the polygon during playback: the camera drifts and pans over 15 seconds, and the zone stays welded to its stretch of road. Every entry event in the replay is evaluated against that per-frame geometry, and you can download the polygon for every frame if you want to check the registration yourself.