Real-time Pothole Detection using Deep Learning

Abstract

The maintenance of infrastructure and the safety of roads are seriously threatened by potholes. Potholes can be quickly located by authorities, allowing for quicker repairs and safer commuter routes. This study uses the YOLO (You Only Look Once) version 5 deep learning model to demonstrate a real-time pothole detection system. Compared to traditional pothole detection approaches, deep learning model detection uses less human power which will contribute to less error too. The suggested method makes use of YOLOv5's high accuracy and efficiency to find potholes in recorded or live video streams. The model detection achieves an mAP (mean average precision) of 76% with 1200 annotated public datasets trained and deployed on the YOLOv5 medium-sized model. Data collection, preprocessing, model training, and inference are all part of the system's end-to-end pipeline. A diverse dataset of road images annotated with pothole locations is used to train the YOLOv5 model. High detection accuracy and real-time performance are achieved by the proposed approach, as shown by experimental results. Under various lighting conditions, weather conditions, and road types, the system's robustness is assessed. The results show that the real-time pothole detection function can be run using a Raspberry Pi but with 2-3 FPS (frame per second) only. With the current FPS, difficulty may occur for real[1]world deployments in high-speed scenarios and need optimization to boost edge computing performance.