A Fused Deep Learning Expert System for Road Damage Detection and Size Analysis Using YOLO9tr and Depth Estimation

This paper presents an innovative fusion of deep learning models for automated road damage detection and measurement. By combining the YOLO9tr object detection algorithm with a grayscale-based depth estimation technique, our system addresses key limitations in traditional 2D assessment methods. The YOLO9tr component enables real-time detection of various road damages, including potholes, cracks, and surface deterioration, while our depth estimation module extracts dimensional information from standard grayscale images using a novel calibration approach. Through extensive testing on diverse road conditions, our system achieved a Mean Absolute Percentage Error (MAPE) of 19.4% in damage size estimation, with consistent performance across various damage types. The system's ability to operate with standard imaging equipment, coupled with its comprehensive calibration methodology spanning 3.5 to 7.0 meters, makes it a practical solution for large-scale infrastructure monitoring. Our approach demonstrates that accurate road damage assessment can be achieved without specialized depth sensors, contributing significantly to the field of automated transportation infrastructure maintenance.