Hybrid Pothole Detection Using YOLO and Contour Based method

Potholes are an important driver of vehicle pare loss and road crashes;thus effective vehicle
pothole detection systems are needed to improve road safety. Traditional pothole detection
can be inefficient, time-consuming, and can use manual-based inspections. Furthermore, with
advances in computer vision techniques, automated and scalable pothole detection systems
can be developed. This project demonstrated a hybrid approach to pothole detection by using
YOLO-based object detection and contour-based post-processing to localize potholes. This
project utilized a custom dataset of road surface images that were annotated in the YOLO
format. The YOLO model detects a bounding box of a potential pothole and contouring
improves the area-specific boundaries within YOLO bounding boxes. The hybrid model,
demonstrating a high number of correctly identified potholes with a lack of false positives.
Nevertheless, there were still some challenges for pothole detection including shadows, water
patches, and cluttered backgrounds that impacts the robustness of the pot hole detection.
Overall, this hybrid-based approach has good potential for real-time deployment in an
intelligent driver-assistance system and can be developed with lane masking, depth filtering,
and improved datasets.