R&D and production
Road Damage Detection — Production Fleet
2025 – Present
From a publicly funded R&D prototype to a production fleet system deployed to a metropolitan municipality, with a 10-vehicle rollout in progress.
- 87%detection accuracy
- 10vehicles in rollout
Publicly funded R&D project · Jun 2025 – Sep 2025
- Led dataset diversification and multi-camera-position data collection (windshield, rear window, roof, trunk), raising model accuracy to 87%.
- Expanded the class taxonomy from 2 to 9 classes and migrated the detection engine from YOLOv4 to YOLOv8.
- Integrated multi-object tracking to eliminate duplicate detections.
- Presented the final on-device Jetson demo at the funding agency’s project review.
Production fleet system · Apr 2026 – Present
- Re-architected the prototype into a production-grade fleet system deployed to a metropolitan municipality (10-vehicle rollout, field installation ongoing).
- Rebuilt the Jetson service architecture to eliminate unnecessary reboots.
- Added system health monitoring and an OTA mechanism for remote model and software updates.
Stack
- Python
- Ultralytics YOLO
- OpenCV
- ByteTrack
- TensorRT
- Jetson Orin Nano
- MongoDB
- GPS/GNSS
- REST API
- LTE
- OTA