Road condition classification from CCTV images using machine learning

Abstract

Understanding and categorizing road conditions is crucial for driver safety and road maintenance. This research explores practical approaches to classify road conditions using images from CCTV stations. Two classification challenges are addressed: distinguishing between snowy and non-snowy conditions and between snowy, wet, and dry conditions. The thesis evaluates various machine learning methods for road condition classification on multiple CCTV stations, including established and novel approaches. Established methods involve feature extraction through texture analysis and finetuning convolutional neural networks and vision transformers. Novel contributions include training an image segmentation model for road segmentation and utilizing persistent homology for feature extraction. Notably, this thesis sets itself apart by separating data into training and test sets based on CCTV stations. This is important to evaluate the methods’ and models’ abilities to generalize to new CCTV stations. The best-performing model, a fine-tuned vision transformer, achieved accuracies of 87.9% and 75.3% for classifying snow/no snow and snow/wet/dry, respectively. These results underscore the complexity of the classification problem and highlight the effectiveness of deep learning models for large-scale road condition classification based on images.