Computer Vision and Machine Learning in Sustainable Mobility: The Case of Road Surface Defects

Road maintenance has traditionally been a time consuming, expensive, and manual process. Timely maintenance of roads helps in lowering rehabilitation costs, accidents, environmental pollution, while facilitating increased connectivity, trade, and growth. Easily acquirable front-view scene images are seen to be used lately for infrastructure management and road maintenance as they provide quicker, low-cost, and flexible solutions. Such scene images can easily be acquired using standard commodity cameras. In this dissertation, machine learning based approaches have been developed to analyze front-view scene images for detecting cracks automatically on road surfaces across different locations and under various conditions. This work thus contributes toward automated approaches to detect different kinds of cracks on road surfaces, thereby proposing a low-cost solution to road maintenance practices. As a result, different components are developed in this work which are sketched together to form a Decision Support System for the task of crack detection. In this study primarily three algorithmic approaches have been developed. Firstly, an unsupervised graph-based hierarchical clustering technique for road area segmentation has been developed, thus helping in detecting the road area in scene images. Secondly, a classifier and superpixel based supervised learning approach consisting of systematically identifying relevant features for detecting superpixels containing cracks has been developed. Thirdly, an unsupervised learning approach consisting of Gamma Mixture Fuzzy Model based clustering technique and keypoint matching mechanisms have been designed in this work for detecting which road pixels are crack pixels in images. Finally, this study integrates the findings and approaches to propose a Decision Support System for crack detection on road surfaces of easily acquirable front-view scene images. Evaluations performed on an experimentally collected diverse front-view scene