Event detection
Given a video, return every event as [start_sec, end_sec, label].
Scored by macro F1 over temporal IoU thresholds 0.3 / 0.5 / 0.7.
WIUT HACKATHON 2026 // COMPUTER VISION
A fixed road camera, one detector, one tracker and a rule engine that understands the intersection: every violation reported as a precise time segment, and an alarm raised before a crash begins.
event classes
false alarms in 18 min
· 29.97 fps input
processing time vs. video length (limit 3×)
PROBLEM
Operators cannot watch every feed, and by the time a crash is reported the footage has already scrolled past. The task splits into two halves.
Given a video, return every event as [start_sec, end_sec, label].
Scored by macro F1 over temporal IoU thresholds 0.3 / 0.5 / 0.7.
At every frame, using only past frames, output the probability that an accident starts within the next 5 seconds.
Strictly causal: no future frames, no lookahead.
PIPELINE // FROM PIXELS TO EVENTS
BRANCH FROM TRACKING // ACCIDENT ANTICIPATION
EXPLORATORY ANALYSIS
18 minutes
4 sample videos, 4K at 29.97 fps, H.264 10-bit 4:2:2 — CPU-only decoding.
noon
noon
evening
dusk

Vehicles crossing the stop line per minute, min to max. On red the rate drops to a handful, so a crossing on red stands out clearly.
up to 3%
The camera shifts up to 3% of frame width between recordings, so the scene map is re-aligned per video.
Where road users actually move: track points per cell on a 48×27 grid, measured from our tracks on all 4 sample videos.
Measured from our tracks on all 4 sample videos. The last minute of each video is partial.
RESULTS ON SAMPLE VIDEOS
READ THIS BEFORE THE NUMBERS
Labels were collected by reviewing the candidates our rules proposed (only C3905 was watched end to end). Events that every version of the rules missed never enter the metric, so recall is an upper bound. On C3905, watching in full revealed 1-2 missed events out of ~10, which puts the honest held-out score closer to 0.8 than to 0.9.
Score A on dev videos (C3896 + C3905)
Held-out videos (C3897 + C3902)
0.901
Score A · dev 0.881
Why higher than dev: On the held-out videos three of six classes have only 1-2 events each and score 1.0, while the dev average is pulled down by jaywalking boundaries.
Scores on our own labels, not the hidden test set.
Score A 0.881 · 20 labelled · 21 predicted
Not scored: accident, near_miss and congestion do not occur in the samples and are not predicted, so they do not enter these scores.
A pedestrian is on the carriageway outside a crossing.
Example: C3896 · 01:52.5-01:57.8 (5.3 s)
28 detected in 18 min
A vehicle drives through a crossing while a pedestrian is on it or stepping onto it.
Example: C3896 · 03:43.2-03:44.0 (0.8 s)
10 detected in 18 min
A turn from the wrong lane or in a prohibited direction.
Example: C3896 · 00:49.3-00:54.8 (5.5 s)
5 detected in 18 min
A vehicle stops past the stop line on red without entering the intersection.
Example: C3896 · 01:06.1-01:42.2 (36.0 s)
4 detected in 18 min
A vehicle crosses the stop line while its signal is red.
Example: C3896 · 01:19.0-01:22.3 (3.3 s)
1 detected in 18 min
A lane change or manoeuvre across a solid marking.
Example: C3896 · 01:10.4-01:12.0 (1.6 s)
1 detected in 18 min
A vehicle stands still on the carriageway for 10 s or more, and not as part of a queue at a signal.
Example: C3897 · 02:25.0-02:50.6 (25.6 s)
1 detected in 18 min
A vehicle moves against the traffic direction of its lane, including driving in the oncoming lane.
Example: C3902 · 00:39.1-00:41.1 (2.0 s)
1 detected in 18 min
Where the system is still wrong, and why. We would rather show this than hide it.
TECHNICAL REPORT
TEAM
Team captain · Frontend, product & analytics
WHO DID WHAT
PREVIOUS PROJECTS
Real Holat, a civic-tech platform for checking how public money is spent on infrastructure (3rd place and a 40M UZS investment grant, Real Holat Hackathon 2026); 2nd place, Uzum AdTech Hackathon 2026; 3rd place, National Transport Hackathon 2026 (Safe Layer, a road-quality map for micromobility).
Backend & infrastructure
WHO DID WHAT
PREVIOUS PROJECTS
Yoshlar Radar, a GovTech youth-employment platform (NEXUS30 hackathon).