A Pothole Can Be Seen with Two Eyes: An Ensemble Approach to Pothole Detection

Abstract

The issue of potholes and bad road quality is faced by nearly every developing country. These conditions often lead to accidents and traffic and cause the general public inconvenience. Significant research work has been performed in the field of effective pothole and road quality detection but due to several underlying issues like relying on only computer vision or sensor-based modeling, or requiring specialized equipment like LiDAR, implementing these solutions in the real world has never been feasible, especially for developing countries where a reliable system is required which works under any condition and more importantly, can be implemented with minimal resources and changes to the existing infrastructure. The proposed system is an end-to-end robust architecture that utilizes mobile devices mounted on government vehicles for data collection and features an innovative ensemble of a YOLOS and a YOLOv8 model resulting in a 97.34% mAP@0.50 score for pothole detection, in conjunction with state-of-the-art sensor based pothole detection and road quality detection models resulting in impressive accuracies of 98.5% and 95.4% respectively. This two-forked approach to pothole and road quality detection combined with innovative connected applications like an analytics dashboard for the government and a navigation application for the consumer make this entire system highly relevant while being scalable, reliable, and easily deployable using the existing infrastructure in developing countries.