| dc.description.abstract | Reducing crash counts on saturated road networks is one of the most significant benefits behind the introduction of Autonomous Vehicle (AV) technology. To date, many researchers have studied how AVs maneuver in different traffic situations, but less attention has been paid to the car-following scenarios between AVs and human drivers. A mismatch in the braking and accelerating decisions in this car-following scenario can lead to rear-end near-crashes and therefore needs to be studied. This thesis aims to investigate the driving behavior of human-drivers that follow a designated AV leader in a car-following situation and compare the results with a scenario when the leader is a human-like driver. In this study, speed trajectory data was collected from 48 participants using a driving simulator. To estimate the near-crash risk between the participants and the leading vehicle, critical thresholds of six Surrogate Safety Measures (SSMs): Time to Collision (TTC), Inverse Time to Collision (ITTC), Modified Time to Collision (MTTC), Deceleration Rate to Avoid Crash (DRAC), critical jerk and Warning Index (WI), were used. The potential near-crash events and the safe driving events were classified using a random forest algorithm after performing oversampling and undersampling techniques. The results from the two-sample t-tests indicated a significant difference between the overall deceleration rates, braking speeds, and acceleration rates of the participants and the designated AV leader. However, no such difference was found between the participants and the human-like leader while braking and accelerating at stop-controlled intersections. Out of six SSMs, MTTC detected near-crash events 10 seconds before their actual occurrence at a range of 11.93 m with 83% accuracy. The surrogate measures identified a higher number of near-crash (high risk) events when the participants followed the designated AV and made braking maneuvers at the stop-controlled intersections. Based on the number of near-crash (high risk) events, the designated AV's C3.25 speed profile (with the maximum deceleration rate of 3.25 m/s2 ) posed the highest crash risk to the participants in the following vehicle. For potential near-crash events classification, a random forest classifier based on undersampled data achieved the highest average accuracy rate of 92.2%. The deceleration rates of the designated AV had the highest impact on the near-crashes between the AV and the participants. However, shorter clearances during the braking maneuvers at intersections significantly affected the near-crashes between the human-like leader and the participants in the following vehicle. | en |
