motc/road-incident-detector
Road Incident Detector
mAP@0.5
0.87
Incident recall
0.90
False alarms / camera / day
1.4
Median detection delay
8.2 s
About this model
Real-time detector that flags collisions, stopped vehicles, debris, and pedestrians-on-carriageway from roadside camera streams. Emits geo-tagged incident events with confidence and a short evidence clip reference for operator confirmation.
Intended use
Automated incident detection feeding traffic-management-centre queues; every alert is designed for human operator confirmation before response dispatch.
Training data lineage
Trained on annotated frames from the traffic-sensor-feeds camera network covering highway and urban arterial scenes in varied light and dust conditions.
Limitations & bias notes
False positives rise at dawn and dusk from long shadows misread as stopped vehicles, and heavy dust storms can suppress detections entirely. Cameras with non-standard mounting heights need per-site threshold tuning.
Evaluation metrics
| mAP@0.5 | 0.87 |
| Incident recall | 0.90 |
| False alarms / camera / day | 1.4 |
| Median detection delay | 8.2 s |
Try it
Live sandboxexpressway_cam_14.jpg
fixtureFeedback
Owning entity
- Updated
- 2026-06-19
- Latest version
- 1.4.0
- License
- Open
- Access
- Open
Trained on
traffic-sensor-feeds →