OpenCV AI Competition 2026 · OpenCV 5 + AWS

Floorsight

Measure forklift risk from existing CCTV. See floor distances, check a predicted close pass, and review the evidence before sending an alert.

Research prototype. Real factory clips are processed offline. Slow and stop commands control only the simulated forklift. No real truck integration or field safety validation is claimed. This page contains recorded demonstrations.

Recorded results · 24 October 2026

2 → 0 contacts

In-path contacts in one simulated scenario. Travel fell from 80 m to 27 m.

3 reviewed warnings

Final AWS run: three approvals and three SNS alerts. The recorded reviewer was scripted.

Measured in metres

Floor geometry accounts for camera perspective. Distances still depend on calibration and detection quality.

New controlled agent checks · 27 October: false exit alerts fell from 8 to 0, with all 8 visible hazards retained. Floor-warning rejection recommendations fell from 16 hazardous cases to 0. These synthetic results and the current policy are explained in the new submission video. Protocol, limits and traces.

Control comparison · Real clips · Privacy checks · Earlier work

The demo

Floorsight submission video · 27 October (3 min 41 s)The problem and solution, recorded app review, current agent results, simulated control, OpenCV 5, AWS architecture, COOL measurements, deployment code and limitations. Historical app footage is labelled; narrated with Amazon Polly, with captions. download

Simulation: what changes when control is on

Four rendered ceiling cameras feed the simulator's OpenCV perception pipeline; the simulated forklift obeys Floorsight's slow/stop commands. Results: in-path contacts 2 of 2 without the interlock, 0 of 2 with it (closest 0.70 m and 0.65 m); 27 m driven in 40 s instead of 80 m.

Interlock OFF (40 s)Floorsight sees the people, nothing brakes, and the forklift reaches both people who step into its path. download
Interlock ON (40 s)Same scene: the forklift slows, stops, waits until each person is seen clear, then drives on. download
The app on the simulated warehouse (1 min 58 s)The Floorsight app itself on the four simulated cameras, issuing AUTO-SLOW, AUTO-STOP and RESUME. download

The app on the real factory clips

Recording correction: the old app resumes at 40.4 s in clip-dev-002. The 27 October fix keeps it stopped until the person clears the path. Replay evidence.

The app on AWS, six real clips (3 min 7 s)Floor map, AUTO-SLOW / AUTO-STOP / RESUME banners (recorded video, so the command is logged, not applied), the agent's one-line recommendation, approve/reject, alerts to the floor lead. download

Real factory clips, measured on the floor (final run, 24 Oct)

Smoothed movement and predicted risk areas drawn on each clip. These are processed copies with detected heads and faces pixelated.

clip-dev-0022 warnings; in the app: slow, stop, resume. download
clip-dev-0031 warning as the forklift pulls away from a stop; in the app: auto-slow. download
clip-dev-004Closest 2.83 m on the floor: no warning (on screen it looks like contact). download
screen-002Our hardest clip: closest 1.55 m, no warning once turn-rate noise is filtered. download
screen-032 download
screen-033 download
Smoothing: raw against Kalman-filteredSame clip, same detections. Left: raw outlines. Right: Kalman-filtered tracks. download

Privacy filtering and failure checks

Filtered clips used for display. A filter exception withholds the frame; undetected heads can still be missed. The development comparison covered 98.0% of reference head regions.

clip-dev-002, filtered download
clip-dev-003, filtered download
clip-dev-004, filtered download
screen-002, filtered download
screen-032, filtered download
screen-033, filtered download
Fail-closed test (12 frames)On frames 4 and 8 the face filter is forced to fail: those frames come out black and marked withheld. Use the player's playback speed or pause to step through. download
Fail-closed test, filter output only download
Earlier videos (superseded)

24 October demo (4:22): preserved for reference; superseded by the 27 October submission video above.

Superseded: some numbers in these were corrected later (for example, screen-002 no longer warns once the turn-rate noise is filtered). The final historical numbers are in the demo and report.

17 Oct: simulated camera 1 (3 min)The first simulated scenario, rendered. download
17 Oct: simulated camera 2 (3 min) download
17 Oct: simulated camera 3 (3 min) download
17 Oct: simulated camera 4 (3 min) download
17 Oct: multi-camera tracking in the simulator download
17 Oct: region history in the simulator download
19 Oct: clip-dev-001, earlier detector stack download
19 Oct: clip-dev-002, earlier detector stack download
19 Oct: clip-dev-004, earlier detector stack download
20 Oct: first danger zones, clip-dev-004 download
20 Oct: first danger zones, screen-002 download
20 Oct: the swerve, half speedWith the earlier detectors the outline swerved in the last 40 ms and the warning came 0.04 s too late. download
21 Oct: first app run on the real clips download
24 Oct: demo draft 1 download
24 Oct: demo draft 2 download