Computer Vision
Roadside Driver-Behaviour Analytics
Identifies a vehicle (body type, plate, colour) and detects driver-caused violations from a single night-time pass through the windshield. Built on one rule: nothing that cannot be proven is reported. Result: zero false positives.
- My role
- AI / computer-vision engineer
- Context
- TEKNOFEST 2026 · 5G & AI Smart Road Safety · Team BiDatalar
- Stack
- YOLO11 · Pose estimation · EasyOCR · OpenCV
1.00
precision
0
false positives
F1 0.77
overall
The problem
A roadside camera sees a car for a few seconds, at night, through the windshield. From that single pass the system has to identify the vehicle (body type, plate, colour) and detect driver-caused road-safety violations: phone use, smoking, passengers, objects on the dashboard, slalom driving. Built for TEKNOFEST 2026 5G & AI Smart Road Safety with team BiDatalar.
One rule guided every design decision:
Nothing that cannot be proven is reported. A false detection costs more than a missed one.
Architecture
Three design decisions worth explaining
1. Driver action: geometry instead of a trained classifier. The trained classifier said smoking with ~99% confidence on all three test videos, and produced seatbelt violation at ~1.0 confidence when the driver left the frame. It was confidently wrong on out-of-distribution input. Instead, the wrist is placed in a nose-centred coordinate frame normalised by face scale:
| lateral (x) | vertical (y) | meaning | |
|---|---|---|---|
| phone | 0.83 | 0.74 | hand beside the head, at ear level |
| cigarette | 0.02 | 1.14 | hand in front of the face, at mouth level |
| no action | 0.06 | 1.57 | hands on the wheel, far below the face |
Physically motivated, explainable and robust to domain shift. If the driver’s face cannot be read, no action is reported.
2. Slalom: detrend the trajectory. As a car approaches and passes, its centre naturally traces a big arc that raw oscillation mistook for slalom. A degree-2 polynomial is fitted and removed, and the remaining lateral RMS is measured in vehicle-width units: straight driving 0.035–0.046, weaving 0.091.
3. Cabin mascot: synthetic domain adaptation. In the public dataset the mascot fills 27% of the image; in the real scene it is a thousandth of the cabin. Crops were composited into real in-cabin scenes as small, dark, blurred, soft-edged objects to create 2,500 synthetic samples, plus a geometric guard against brake lights.
Results
Vehicle info: colour 3/3, body type 2/3, plate 2/3. On the third plate the system deliberately abstained because of the angled shot, returning null instead of inventing one.
How it got there
| Issue | Before | After |
|---|---|---|
| Driver action | 7 of 10 detections wrong (P ≈ 0.30), contradictory labels | P = 1.00, R = 1.00 |
| Cabin mascot | never detected even when clearly visible (0.10 conf.) | detected (up to 0.82) |
| Passengers | the driver counted as “front passenger” | driver excluded |
| Body type | a C-SUV called “sedan” at 99.8% | 2/3 correct |
| Slalom | disabled (false positives) | detected |
Delivery
An offline Docker image on CUDA 12.1 with embedded weights, a normalisation layer (CLAHE + gamma) and adaptive frame stride. It writes a JSON result validated against the committee’s schema. It is always written, even when every head abstains.
