Investigating the Contributing Factors to AV-Road User Conflicts: A Data-Driven Approach

Abstract

This thesis aims to develop a solid understanding of the factors that impact the safety level of Autonomous vehicles (AVs). To that end, two studies were conducted. The two studies utilized the Woven prediction and validation dataset to extract AV-road user conflicts. The extracted conflicts were classified by type, location (road segment and intersection), and severity (severe and non-severe). In the first study, a copula-based modelling approach was used to investigate the impact of a wide range of factors on the frequency of conflicts. For road segments, the results showed that the hourly rates of conflicts increased on large, major, and divided roads but decreased on roads with bike lanes. The posted speed, presence of bus stops, presence of on-street parking, and the density of access points were identified as the key factors that impact conflict occurrence. For intersections, the results show that posted speed, presence of median, and presence of pedestrian crossing are the key contributing factors to the hourly conflict rates, while the presence of traffic signals and bike lanes was shown to be associated with lower rates of severe AV conflicts. The second study investigated the impact of various variables on the severity level of AV conflicts, using the Logistic regression and ML Multilayer Perceptron modelling approaches. Results showed that the average speed of vehicles around the AV, road user volume, land use, and road type are the key variables that determine conflict severity in road segments. At intersections, traffic control devices, road user volume, the presence of protected left turn phases, and the average speed of vehicles around the AV were the most impactful factors. For both road segments and intersections, pedestrian volume was directly associated with higher conflict severity, while the presence of bus stops, bike lanes, and parking lanes led to less severe conflicts.