Computer Vision-Based Real-Time Pothole Detection and Severity Assessment System for Urban Road Maintenance

Deteriorating Road Surfaces Remain A Major Contributor To Traffic Accidents, Vehicle Damage, And Rising Municipal Maintenance Costs In Fast-Urbanizing Cities Of South-East Asia. This Paper Presents The Design, Deployment, And Field Evaluation Of A Low-Cost, Real-Time Pothole Detection And Severity Assessment System Based On A Yolov8 Deep-Learning Model Running On An Edge Computing Platform (Nvidia Jetson Nano). Dash-Cam Video, Gps Coordinates, And Inertial Measurement Data Are Fused On-Board To Detect Potholes, Classify Their Severity Into Four Categories (Low, Medium, High, Critical), And Stream Geo-Tagged Reports To A Centralized Municipal Dashboard Via 4g. A Custom Dataset Of 12,400 Annotated Images Was Collected Across 8 Districts Of Johor Bahru And Used To Train And Validate The Model. Field Trials Over 480 Km Of Urban Roads Achieved A Mean Average Precision (Map@0.5) Of 0.93, An Inference Rate Of 27 Fps On The Edge Device, And A Severity Classification Accuracy Of 91.4%. The Proposed System Reduced Manual Road-Survey Effort By Approximately 78% And Enabled Prioritized Repair Scheduling, Demonstrating A Practical Pathway Toward Data-Driven Smart Road Maintenance For Medium-Sized Municipalities.