Enhancing Road Anomaly Detection With Dynamic Cropping System: a YOLOv8 Integrated Approach

Abstract

Efficient and accurate detection of urban road anomalies such as potholes, manhole covers, and speed bumps is crucial for enhancing urban infrastructure and ensuring road safety. However, detecting these small-scale features using machine learning is significantly challenged by the high prevalence of negative data and the complex urban backgrounds in images. This study introduces an innovative approach utilizing a Dynamic Cropping System (DCS) in conjunction with the YOLOv8 convolutional neural network model to refine the detection of these road anomalies. The DCS method enhances detection accuracy by employing a YOLOv8- based model to identify a nd i solate r oad s urfaces w ithin i mages, t hereby minimizing irrelevant background information through targeted cropping.