Analysis of Urban Intersections Using Computer Vision and Machine Learning

The crashes, injuries, and fatalities that occur at roadway intersections year after year are a tremendous source of economic and societal loss. Intersections are the most complex components of an urban roadway network, consisting of conflicting movements, varying mode choice, and high traffic volume. While there has been no rest academically and at the professional level to address these issues, the multitude of variables, both observable and unknown, that contribute to danger at an intersection makes comprehensive analysis difficult at a local scale, and unfeasible at the macro level. To this end we propose a novel approach to analyze urban intersections, leveraging the use of machine learning, artificial intelligence, and computer vision on a readily available source of data, aerial intersection images. The research consists of applying both unsupervised and supervised machine learning techniques on a custom dataset synthesized from multiple data sources. Geospatial data located in the project area was used to obtain a representative sample of urbanized intersections regardless of typical research considerations such as geometric design, approach grade, control-type, or capacity. These intersections were then quantified by combining aggregated crash records and historical traffic data to provide base metrics for each intersection location that can be used to relate instances in the dataset with one another and establish base metrics. Exploratory data analysis was then applied to this synthesized dataset to uncover patterns and relationships in the dataset and to obtain key statistical information that informed the development of the subsequent machine learning models. An unsupervised machine learning model was first created to assess the entirety of the dataset, and to cluster intersections based on their physical similarities. From there, statistical analysis based on recorded intersection crashes and traffic data were performed to identify groups of intersections in the dataset that were more likely to have crashes occur. Supervised machine learning techniques were then used to reframe intersection safety analysis in the form of an image classification model, whose objective is to predict an assigned safety label based on the previously obtained quantitative values. This thesis presents a broad application of machine learning techniques and coupled with crash data and traffic volume data, reframes the complex task of intersection safety analysis as a machine learning task that an artificial intelligence can be trained to learn. The results of this research is a scalable and efficient unsupervised dimensionality reduction model that can be used to perform a global assessment of intersections that can group intersections in terms of there physical similarities, and a set of supervised, binary image classification models that can predict an assigned safety label with an average accuracy of 81.0%. The models developed for these tasks show the successful application of machine learning technologies in the context of intersection safety analysis at a scale unseen in traditional and contemporary studies.