Enhanced Autonomous Driving

Abstract

The process of repairing potholes entails significant financial and temporal resources. This article presents cutting-edge approaches to pothole detection employing convolutional neural network techniques, focusing exclusively on RGB input images. The primary objective of this research is to assess the performance of three iterations of the You Only Look Once model, namely, YOLOv8 Nano, Small, and Medium, for pothole detection. Additionally, the proposal suggests a system to alert drivers about identified potholes and recommend alternative routes. The design of the solution enables it to function effectively in real-time in various lighting conditions, including low-light scenarios. Evaluation metrics include inference speed and detection accuracy. We merged diverse datasets from multiple sources to facilitate model training and accelerate convergence. The mean average precision at IoU threshold 0.5 (mAP@0.5) for YOLOv8 n, s, and m models is reported as 66%, 71%, and 51%, respectively.