WIUT Hackathon 2026 · Computer Vision track
Traffic events from a fixed road camera, as time segments you can click.
We detect and track every road user, align a hand-drawn map of the junction to each clip, and turn trajectories into events such as jaywalking, red-light running and congestion.
- sample clips analysed
- 4
- event classes detected
- 7 / 14
- camera resolution and fps
- 4K · 29.97
- scene-map alignment residual
- < 0.5 px
How it works
One learned component, everything after it deterministic. Read the full approach.
- MP4 video4K, 29.97 fps, fixed camera
- Frame samplingRule-basedDecode once, analyse a subset of frames
- YOLO detectorLearnedFine-tuned, 8 classes incl. red/green signal heads
- ByteTrackRule-basedTrack ids and trajectories per road user
- Scene alignmentRule-basedSIFT + RANSAC homography maps scene.json onto the clip
- Event rulesRule-basedPer-class geometry and timing rules on tracks and zones
- Segment mergingRule-basedMerge fragments, drop blips, no same-class overlap
- Events[start_sec, end_sec, label]
Blue border = learned component. Everything else is deterministic code whose thresholds are listed below.
Sample videos
Playback, event timeline and EDA for every clip the organizers gave us.

C3896
5 min 40 sPedestrian on roadway · Not yielding to a pedestrian · Illegal turn · Solid line crossing · Stop-line violation · Congestion
Open results
C3897
5 min 18 sIllegal U-turn · Pedestrian on roadway · Not yielding to a pedestrian · Solid line crossing · Stop-line violation
Open results
C3902
5 min 18 sRed-light running · Illegal U-turn · Pedestrian on roadway · Not yielding to a pedestrian · Illegal turn · Solid line crossing · Stop-line violation
Open results
C3905
2 min 8 sNear miss · Pedestrian on roadway · Not yielding to a pedestrian · Solid line crossing · Stop-line violation
Open results