Risk Analysis Using Artificial Intelligence Algorithms to Prevent Collisions on Roadway Segments

This thesis focused on improving the risk analysis algorithms used in collision avoidance systems (CASs) designed to reduce the risk of three types of collision on roadway segments: animal-to-vehicle collisions, pedestrian-to-vehicle collisions, and pedestrian-to-pedestrian collisions. Currently available CASs use only one input indicator. This approach is limited as the CASs: apply a simple risk analysis algorithm based on a fixed threshold to identify risky situations; cannot simultaneously capture a variety of important collision contributing factors; and cannot combine multiple contributing factors into a single composite risk indicator. The goal of this thesis was to use artificial intelligence algorithms to create a composite risk indicator based on a combination of various input indicators. The thesis goal was achieved through four objectives:

1. Develop a fuzzy rule-based algorithm for a next generation roadside animal detection system;
2. Develop a fuzzy rule-based algorithm for a smart protection system to reduce the number of collisions with police officers on duty on the roadway;
3. Develop a semi-supervised machine learning algorithm for a smart protection system to reduce the number of collisions with police officers on duty on the roadway; and
4. Develop a risk analysis approach to evaluate physical distancing on urban sidewalks. Improvement of the existing risk analysis algorithm in objective 1 resulted in capturing driver behavior, animal behavior, and the spatial and temporal interaction between animal and vehicle. It also resulted in differentiating risk for following and leading vehicle and generating no-risk when vehicle passed from animal. Objectives 2 and 3 were part of the same CAS study. Improvement of the existing risk analysis algorithm in both objectives 2 and 3 resulted in capturing pedestrian behavior, driver behavior, the spatial and temporal interaction between pedestrian and vehicle with 94% accuracy when estimating all risk labels, and 88% success when identifying near miss collisions. Objective 4 successfully reflected the role of density and exposure time in the level of physical distancing. It could help decision-makers to select the most appropriate interventions (e.g., sidewalk expansion) for pedestrians to maintain physical distancing.