05 /AI / COMPUTER VISION
TRINETRA.
Watches CCTV and drone video of crowded places like temples, spots when a crowd is turning dangerous, and alerts officers on their phones before it becomes a stampede.
How it works
See the crush before it happens.
Crowd-risk scoring from live video. ResNet-50 detects people and estimates density and motion per frame; an XGBoost model over 80+ engineered features classifies risk as low, moderate or high, and alerts go out over WebSocket. The budget that shapes everything: about 120 ms per frame and under 1.5 s from frame to alert.
The hard parts
- Holding per-frame inference near 120 ms so alerts stay live.
- Turning raw counts and motion into features a classifier can rank risk on.
- Getting an alert from the model to an officer's phone in under 1.5 seconds.
Results
- 85% detection accuracy and 75% risk-prediction accuracy, as reported in the project README.
- Live heatmaps and alert history on the dashboard and the mobile app.
Stack
Built for TechJam 2.0 (Merit prize)
Reach the end of the file to complete this operation
