PariVision

WIUT Hackathon 2026, CV track

Traffic events at one Tashkent junction, read from a fixed 4K camera.

We are PariVision, three students from Webster University in Tashkent. Our system watches a signalised junction and does two jobs: it lists traffic events as time segments with a class, and it gives a per-frame risk that an accident starts within the next 5 seconds.

Try the demoSee the sample clips
A short loop from C3896 with our boxes, event labels and signal phase drawn on. The camera sits high on a building at one corner of the junction and looks up the avenue.
0.523Dev-set Score Amean F1 over classes, IoU 0.3 / 0.5 / 0.7
3.28minProcessing per minute of 4K videoPart A + Part B, Apple M5 laptop on MPS, time guards offThe harness allows 3x. On a Colab T4 our two detectors take 0.53x; decoding the 4K files on the CPU is the slow part.
8of 14Event classes we emitU-turns found but not submitted; 5 have no rule
18.4minFootage in our dev set4 clips, 11:18:24 to 17:22:21

From frames to event segments

Part B runs next to it: a smaller detector at 10 Hz, only past frames, time-to-collision between road users turned into a probability.

  1. 01SampleEvery 3rd 4K frame, resized to 1920 px
  2. 02Detect and trackYOLO26m, then ByteTrack per object group
  3. 03RegisterHomography into one reference view
  4. 04Read the signalLamp colour on the median head, every 0.1 s
  5. 05Apply rulesTrajectories plus phase give per-actor intervals
  6. 06SegmentsMerged per class, written as [start, end, class]

The full pipeline, with the rule for each class

Where to look