Experimental Evaluation of Pothole Detection and Its Dimension Estimation Using YOLOv8 and Depth Camera for Road Surface Analysis

Pothole detection and dimension estimation are essential to improve the safety and comfort of autonomous vehicles. This paper uses a depth camera to present a new approach for position estimation and pothole detection using the You Only Look Once (YOLO) method. Experimental results on detection applications using real cars show good pothole detection accuracy in both bright and dimly lit road conditions covered by trees with precision. This object detection uses the YOLO version 8 nano model accompanied by coco as pre-trained training. In the training dataset, only one class, namely the pothole dataset, is used. A depth camera from Intel Realsense type D455 will be employed, and the Jetson Orin Nano will subsequently apply the training results. During the field test, data about the position of the pothole coordinate inside the pixel frame, the pothole's width, and its distance from the camera will be displayed, in addition to pothole-detecting objects. Every pothole object found will have its data shown in real time. Validating the width and length of the pothole involves taking actual measurements with a meter. The estimated distance and width of potholes showed good agreement with direct manual measurements with R -squared values above 0.97 and gradients approaching unity.