The first AI system to detect all five two-wheeler traffic violations — no helmet, phone use, triple riding, wheeling, and illegal parking — in a single real-time processing pass on live Indian road footage. Auto-extracts number plates and dispatches e-challans with zero false positives.
All five classes detected simultaneously in a single DRISHTI-VD™ real-time processing — no sequential pipeline. Each class trained on a custom Indian road dataset (5,800 training / 1,210 validation samples).
The key innovation of DRISHTI: the same proprietary AI engine is reused for both violation detection and number plate localisation — eliminating the extra computational cost of a separate detector.
All metrics independently validated on real Indian road footage. Reference hardware: Intel i5-equivalent, 16GB RAM — no GPU required for standard deployment.
| Violation Class | Actual Violations | Correct Detection | Incorrect Alert | Precision (AP) |
|---|---|---|---|---|
| No Wearing Helmet | 1,045 | 1,041 | 13 | 0.987 |
| Phone While Riding | 802 | 788 | 22 | 0.973 |
| Triple Riding | 165 | 165 | 1 | 0.994 |
| Wheeling | 45 | 45 | 0 | 1.000 |
| No Parking | 21 | 19 | 1 | 0.950 |
| Overall Accuracy | — | — | — | 0.981 (98.09%) |
| Stage | Samples | Visible/Readable | Correct | Success Rate |
|---|---|---|---|---|
| Detection | 2,058 | 1,853 | 1,849 | 99.78% |
| Segmentation | 1,849 | 1,845 | 1,843 | 99.89% |
| Recognition (OCR) | 1,843 | 1,843 | 1,842 | 99.95% → Overall: 99.41% |
| Video | FPS | Actual | Detected | Plates | V-Error% |
|---|---|---|---|---|---|
| Video 1 | 24 | 5 | 5 | 5 | 0% |
| Video 2 | 24 | 13 | 11 | 10 | 15.4% |
| Video 3 | 24 | 22 | 20 | 18 | 9.1% |
| Video 4 | 36 | 1 | 1 | 1 | 0% |
| Video 5 | 34 | 3 | 3 | 3 | 0% |
| Video 6 | 44 | 13 | 11 | 11 | 15.4% |
| Video 7 | 43 | 14 | 14 | 13 | 0% |
| Video 8 | 56 | 22 | 18 | 13 | 13.6% |
| Total | — | 93 | 83 | 79 | — |
Real-time operational view for traffic control rooms — multi-camera, all five violation classes, live e-challan status. Data feeds from paper-validated detection pipeline.
Validated against leading competing systems across performance, accuracy, and India-specific deployment criteria.
| Feature / Metric | DRISHTI Patented · Made in India |
Competitor A Sequential Pipeline |
Competitor B No Object Tracking |
Traffipax Traffistar Germany |
Manual Police Traditional |
|---|---|---|---|---|---|
| Detection approach | Single-pass proprietary AI engine | Sequential: AI Detector→Neural Network→Segmentation Model | AI Detector (no tracker) | Speed radar + camera | Human eyes |
| No. of violations detected | 5 simultaneously | 2 (helmet + crosswalk) | 2 (triple + speed) | Speed only | Variable |
| Object tracker (no re-count) | ✓ Proprietary Multi-Object Tracker | ✗ | ✗ Re-counts violations | ✗ | Manual |
| Number plate extraction | ✓ Shared engine — no extra model needed | ✗ | ✗ | Separate ANPR unit | Manual note |
| Detection Accuracy | 98.09% Accuracy | Lower (seq. pipeline) | Lower (no proprietary optimisation modules) | ~90-94% (speed) | ~70-80% |
| Incorrect Alert Rate (live) | 0.00% | Higher | Higher (re-counts) | ~2-5% | ~8-15% |
| Real-time on consumer CPU | ✓ i5-8250U, no GPU | ✗ GPU required | Partial | ✗ Dedicated hardware | ✓ Human |
| Indian number plate formats | ✓ BH, State codes, all RTOs | ✗ | ✗ | ✗ EU plates only | ✓ |
| Indian road scenario training | ✓ 5,800 Indian road images | ✗ | ✗ | ✗ | ✓ |
| Auto e-challan + DB storage | ✓ Full pipeline | ✗ | ✗ | Partial | ✗ |
| Published peer-reviewed results | ✓ Independently Validated | Conference only | Journal (lower tier) | ✗ Commercial | N/A |
| Inference speed (live video) | 20 fps (frame-skip) | Not real-time | Lower FPS | Hardware-specific | Human reaction |
| System | Method | Result |
|---|---|---|
| DRISHTI | DRISHTI Proprietary Engine | Best |
| Competitor A | Sequential Multi-Model Pipeline | Lower — fails multi-violation |
| Competitor B | Single-model, no tracking | Lower — no tracker |
| Metric | DRISHTI | Competing System |
|---|---|---|
| Speed | Superior | Slower |
| Performance | Superior | Lower |
| Duplicate counts | 0 (DRISHTI Tracker™ eliminates re-counts) | Many (no tracker) |
| Extra plate model | None (reuse) | Separate model |
No GPU needed. Validated on an off-the-shelf Intel i5 laptop — dramatically lowering deployment cost for Indian municipal corporations and state traffic police.
The paper's database + ticketing layer connects to the full Government of India digital infrastructure — no foreign software dependency.
City-scale real-time two-wheeler surveillance across 8 active DRISHTI cameras. Each detection on the canvas feeds directly into this map — same vehicle, same timestamp, same track ID.