Application and Implementation of Deep Learning-Based Road Pothole Detection and Segmentation Algorithms

Abstract

Road pit detection is one of the important means to maintain road and traffic safety. Road pit detection includes target detection of pit holes in the video based on the driving recorder, and judging the location and size of road pit holes. This study aims to accurately judge the pit holes on the road, the size of the pits and the pit levels of the pits based on the driving recorder. Provide valuable pit repair advice for road inspectors, thereby reducing their workload and improving patrol efficiency. This study collected more than 2,000 images of real road pits containing complex conditions such as shadows and strong light in many places in Sichuan, and constructed a high-quality labeling data set. In the object detection task, compared YOLOv8 and Mask-R-CNN, select YOLOv8 as the basic network framework to identify potholes in on-board videos and capture images. Then, the three semantic segmentation models of YOLOv8, U-net and Mask-R-CNN were compared. Since Mask-R-CNN can better segment small potholes, it is selected as the segmentation model. The area is estimated and the inner diameter of the pothole is calculated using the mask to evaluate the warning level. After rigorous training and evaluation, the mAP50 of YOLOv8 for pits was 0.88 and the accuracy was 0.89. The frequency weight ratio of Mask-R-CNN in the semantic segmentation part was 0.55 and the ratio of 0.45. Experiments show that the system detection speed reaches 100FPS and the early warning accuracy rate exceeds 90%.