Transportation Safety Analytics with Statistical Machine Learning

Abstract

Transportation safety has been a topic of high priority for a long time, especially for the society’s public service systems, e.g. the freight transportation system and public commuting system. With the development of modern transportation technologies, complex products have been designed and are being used by the public, such as the connected and autonomous vehicles. The potential risks brought by safety uncertainty of the transportation system is of public’s concerns. Traffic accidents are among the top concerns. For example, in the past decade, there are approximately 100,000 traffic accidents in Arizona and 6,000,000 accidents nationwide every year. Some of them may incur very huge loss and some of them were observed to happen very often. Fortunately, with recent progress in Internet of Things (IoT), big data and machine learning technologies, there are new methods to analyze problems existing in the transportation safety analytics. This dissertation investigates several important transportation safety issues. Utilizing multiple sources of data, new methods for analyzing accident severity and frequency are developed with consideration of the unique characteristics of accident data. In the accident severity analysis, a new method is proposed and applied in the imbalanced multi-class classification problem. Significant safety factors are identified and counter measures are proposed for improving safety. In the accident frequency analysis, a new quantitative method of predicting and assessing accident risks is proposed that takes into account of the relationship between travel time reliability and crash frequency. Besides analyzing historic traffic accidents, this dissertation explores the utilization of connected vehicle data for improving safety. Machine learning techniques are used to predict a vehicle’s trajectory so that potential conflicts and crashes can be identified. This is a powerful new approach to analyzing vehicle data, but can also pose a risk to privacy and security. To investigate the risk, several supervised learning techniques are used to re-identify vehicles from vehicle data used for prediction. In a connected vehicle environment, connected vehicles have anonymous identifications. Using partial trajectory information from different intersections, the same connected vehicles are shown to be re-identifiable with high accuracy using machine learning techniques. Different factors that impact on the re-identification accuracy are evaluated within a designed experiment. For the purpose of predicting connected vehicle’s future trajectory, deep learning techniques are applied with neural network structures with considerations of vehicle information, intersection signal and phase information and surrounding dynamic information. The proposed method is validated and compared with the classic object tracking algorithm and the naïve deep learning method.