Multi-Level Pothole Detection Using Yolo v8 Nanosegmentation

The project aims to detect potholes on the road using a vision-based and deep-learning approach. Potholes are a common road hazard, inflicting severe damage to infrastructure and users. Effective pothole detection will help to avoid vehicle damage, decrease safety hazards, maintain smooth traffic flow, and extend the life of road infrastructure. To address the problem, the proposed work focuses on a vision-based deep learning technique for detecting potholes and classifying them into three levels based on size. The CNN model is implemented in the study’s methodology to determine whether a given image contains a pothole or a normal one and is accomplished by combining five distinct datasets to enhance the total number of images while simultaneously enhancing the effectiveness of the model. The total images obtained are 3315 potholes and 3314 normal images. The potholes are further classified as high, low, or medium using the YOLOv8 nanosegmentation model and a random forest classifier model. A single dataset named "Pothole Image Segmentation Dataset" is used to train the model, and to address the unequal distribution of classes, a synthetic minority oversampling method (SMOTE) is used. Multi-level pothole detection is implemented with perfect precision, recall, and an F1 score of 1.00 for low-class, 0.99(near to ideal) for medium-class, and 0.80 for high-class samples. The lower scores for the high class compared to the other classes reveal significant difficulties in correctly detecting high-severity samples. Even under low scores of high-class prediction, the model achieved an accuracy of 0.99.