TOWARDS ROAD DAMAGE DETECTION BY USING YOLOV8

Abstract

Damages on roads (small holes, large holes, and cracks etc. - potholes in general) are one of the main cause to traffic accidents, disaters and vehices’ devastations. Timely and accurate detection of these potholes and then repair them will help to reduce accidents and damages to roads and vehicles together. This also provides good supports for managers to plan road maintenance and repair more appropriately and optimally. However, manually detecting and measuring road surface damage is a tedious, time-consuming task that is not always feasible and timely. In this article, we propose a solution to the problem of detecting damage on road surfaces (potholes) by applying the YOLOv8 deep learning network model on video data obtained from surveillance cameras. Training results on an image dataset of 19074 images show that the proposed system has good training accuracy (mAP@0.5 is 87.5% with the YOLOv8x model after 200 training epochs). Experimental results of the trained model upon real-life road videos also show that the approach using the YOLOv8 network has high accuracy and is applicable in practice.