Pothole Detection Using Deep Learning: A Real-Time and AI-on-the-Edge Perspective

Asphalt pavement distresses are the major concern of underdeveloped and developed nations for the smooth running of daily life commute. Among various pavement failures, numerous research can be found on pothole detection as they are injurious to automobiles and passengers that may turn into an accident. This work is intended to explore the potential of deep learning models and deploy three superlative deep learning models on edge devices for pothole detection. In this work, we have exploited the AI kit (OAK-D) on a single-board computer (Raspberry Pi) as an edge platform for pothole detection. Detailed real-time performance comparison of state-of-the-art deep learning models and object detection frameworks (YOLOv1, YOLOv2, YOLOv3, YOLOv4, Tiny-YOLOv4, YOLOv5, and SSD-mobilenetv2) for pothole detection is presented. The experimentation is performed on an image dataset with pothole in diverse road conditions and illumination variations as well as on real-time video captured through a moving vehicle. The Tiny-YOLOv4, YOLOv4, and YOLOv5 evince the highest mean average precision (mAP) of 80.04%, 85.48%, and 95%, respectively, on the image set, thus proving the strength of the proposed approach for pothole detection and deployed on OAK-D for real-time detection. The study corroborated Tmy-YOLOv4 as the befitted model for real-time pothole detection with 90% detection accuracy and 31.76 FPS.

Automated summary

This study aims to explore the potential of deep learning models for pothole detection on edge devices, and provides a detailed performance comparison of various models. The results demonstrate that Tiny-YOLOv4, YOLOv4, and YOLOv5 achieve high accuracy and mean average precision in pothole detection under different road conditions and illumination variations.