Comparative study of pothole detection using deep learning on smartphone

Abstract

Potholes present a significant problem in many countries, leading to vehicle damage and traffic accidents. These road imperfections pose safety risks and impose economic burdens. Despite existing detection methods using sensors and computer vision deep learning processed on PCs, a gap remains in deploying cost-effective, widely accessible solutions. This study aims to bridge this gap by developing deep learning models optimized for smartphones, reducing costs and enhancing deployment feasibility. We developed multiple models for pothole detection, utilizing transfer learning and Bayesian hyperparameter tuning to optimize detection accuracy and resource efficiency. Our evaluations focused on computationally light models such as YOLOv8 small, YOLOv8-nano, YOLOv7 tiny, and faster R-CNN MobileNetV3. In terms of detection accuracy, YOLOv8 small and YOLOv8 nano stood out, achieving average precisions (AP) of 83.5% and 82.5%, respectively. YOLOv8 nano proved the most efficient, offering high detection accuracy, a file size three times smaller than YOLOv8 small in TFLite format, and the fastest inference time of 0.72 seconds per image. This study highlights the potential of smartphones in urban pothole detection, contributing to improved road maintenance and urban policy.