AI-driven Vision-based Pothole Detection for Improved Road Safety

Cracked and potholed roads frequently cause deadly accidents, posing serious safety risks and significant maintenance expenses. Vehicles hitting potholes can damage road furniture, increase maintenance costs, and leave road users with significant repair expenditures for their vehicles. Drivers feel insecure and uncomfortable when continually monitoring road conditions to avoid potholes, which detracts from their entire driving experience. This project seeks to create a Pothole Detection System that employs Convolutional Neural Network (CNN) algorithms to explore feature extraction approaches for identifying road potholes. The model was trained with CNN algorithms to identify photos as a pothole or normal, and You Only Look Once (YOLO) to detect and estimate pothole areas. Two datasets were joined to create a cohesive dataset with 681 images from the first and 4000 images from the second, for 4681 images. These pictures, divided between potholed and typical roads, were cleaned and resized to 256×256 pixels. The dataset was split into two groups: 70% training and 20% testing. Roboflow Annotate was used to annotate images. Following data preparation, the CNN and YOLO algorithms were created independently. The CNN-YOLO model had an accuracy rate of 92.85%. This project increases road safety, infrastructure upkeep, and traffic flow. The pothole detection technology warns drivers about potential road hazards, decreasing accidents and fatalities. Efficient detection allows for preventive, cost-effective road maintenance, optimises government resource allocation, and enhances the driving experience by lowering car maintenance costs and assuring safer roads.