Final year project · USM · 2025/26
AI-Based Road Hazard Detection
Report road damage from your phone, let AI check it, and send confirmed hazards to the repair crew.
Problem
Road damage is found through slow manual checks and scattered complaints, and every report is verified by hand before repairs start.
Under the hood
- Inference
- Two passes with CLAHE enhancement and a shadow guard on top of raw YOLO output
- Stress test
- 6 tropical conditions: glare, canopy shadow, rain, wet road, haze, low light
- Retraining
- Admin corrections become training data; a new model goes live only if mAP@0.5 improves
- Deploy
- Flask API in Docker on Cloud Run, React app on Firebase Hosting
The repair loop I built
- 01ReportCitizens snap a photo on a mobile-first web app. GPS is read from the photo, no sign-up.
- 02DetectYOLOv8 marks potholes and cracks with a confidence score.
- 03VerifyAdmins confirm or correct on a dashboard with a live hazard map.
- 04RepairContractors get an email link to accept, reject and complete the job.














