Pothole Detection and Estimation of Repair Cost in Bangladeshi Street: AI-based Multiple Case Analysis

Abstract

This study focuses on the real-world application of a state-of-the-art convolution neural network (CNN) based pothole detection model and its practical implications. A multiple-scenario analysis was performed, combining several experiments based on video recorded at different environmental conditions and vehicle speeds. Moreover, the performance of the CNN-based YOLOv4-tiny AI model was compared with an expert human grader (civil engineer). The findings from the comparative analysis suggest that out of five different cases, the AI-based model (69.57-85.00%) outperformed the human evaluator (43.67-80.67%) in four cases, with the highest accuracy of 85%. This indicates the utility of using an AI-based approach to pothole detection, especially in regional areas of developing countries like Bangladesh.