Author: Pei Li

Publisher:

Published: 2021

Total Pages: 0

ISBN-13:

Recently, with the development of connected vehicles and mobile sensing technologies, vehicle-based data become much easier to obtain. However, only few studies have investigated the application of this kind of novel data to real-time traffic safety evaluation. This dissertation aims to conduct a series of real-time traffic safety studies by integrating all kinds of available vehicle-based data sources. First, this dissertation developed a deep learning model for identifying vehicle maneuvers using data from smartphone sensors (i.e., accelerometer and gyroscope). The proposed model was robust and suitable for real-time application as it required less processing of smartphone sensor data compared with the existing studies. Besides, a semi-supervised learning algorithm was proposed to make use of the massive unlabeled sensor data. The proposed algorithm could alleviate the cost of data preparation and improve model transferability. Second, trajectory data from 300 buses were used to develop a real-time crash likelihood prediction model for urban arterials. Results from extensive experiments illustrated the feasibility of using novel vehicle trajectory data to predict real-time crash likelihood. Moreover, to improve the model’s performance, data fusion techniques were proposed to integrated trajectory data from various vehicle types. The proposed data fusion techniques significantly improved the accuracy of crash likelihood prediction in terms of sensitivity and false alarm rate. Third, to improve pedestrian and bicycle safety, different vehicle-based surrogate safety measures, such as hard acceleration, hard deceleration, and long stop, were proposed for evaluating pedestrian and bicycle safety using vehicle trajectory data. In summary, the results from this dissertation can be further applied to real-time safety applications (e.g., real-time crash likelihood prediction and visualization system) in the context of proactive traffic management.