A Novel Framework for Pothole Area Estimation Based on Object Detection and Monocular Metric Depth Estimation

Currently, autonomous driving technology is rapidly developing. Accurate detection and area estimation of potholes are crucial for enhancing the safety of roads. Previous studies typically relied on physical models based on camera angles or LI-DAR data for pothole area estimation, which often suffered from significant errors and limited range capabilities. To address these issues, a novel framework for pothole detection and area estimation is proposed. Initially, potholes are detected using the high-precision yet lightweight object detection network YOLOv5n-p6; subsequently, the metric depth of pothole keypoints is estimated via the monocular metric depth estimation model ZoeDepth; finally, a pinhole camera model is utilized to compute the area of potholes. Experimental results demonstrate that established pothole detection model maintains high accuracy while achieving model lightweightness, and the proposed area estimation model provides predictions that closely match the actual pothole areas. This research offers a new methodology for pothole detection and area estimation, potentially improving road safety in autonomous driving.