Date on Master's Thesis/Doctoral Dissertation

12-2021

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Civil and Environmental Engineering

Degree Program

Civil Engineering, PhD

Committee Chair

Li, Zhixia

Committee Co-Chair (if applicable)

Sun, Zhihui

Committee Member

Sun, Zhihui

Committee Member

Mohsen, J. P.

Committee Member

Desoky, Ahmed

Author's Keywords

Driver distraction; automated vehicle; quantification; artificial intelligence; traffic safety; human-vehicle interaction

Abstract

Automated Vehicle (AV) technology expects to enhance driving safety by eliminating human errors. However, driver distraction still exists under automated driving. The Society of Automotive Engineers (SAE) has defined six levels of driving automation from Level 0~5. Until achieving Level 5, human drivers are still needed. Therefore, the Human-Vehicle Interaction (HVI) necessarily diverts a driver’s attention away from driving. Existing research mainly focused on quantifying distraction in human-operated vehicles rather than in the AV environment. It causes a lack of knowledge on how AV distraction can be detected, quantified, and understood. Moreover, existing research in exploring AV distraction has mainly pre-defined distraction as a binary outcome and investigated the patterns that contribute to distraction from multiple perspectives. However, the magnitude of AV distraction is not accurately quantified. Moreover, past studies in quantifying distraction have mainly used wearable sensors’ data. In reality, it is not realistic for drivers to wear these sensors whenever they drive. Hence, a research motivation is to develop a surrogate model that can replace the wearable device-based data to predict AV distraction. From the safety perspective, there lacks a comprehensive understanding of how AV distraction impacts safety. Furthermore, a solution is needed for safely offsetting the impact of distracted driving. In this context, this research aims to (1) improve the existing methods in quantifying Human-Vehicle Interaction-induced (HVI-induced) driver distraction under automated driving; (2) develop a surrogate driver distraction prediction model without using wearable sensor data; (3) quantitatively reveal the dynamic nature of safety benefits and collision hazards of HVI-induced visual and cognitive distractions under automated driving by mathematically formulating the interrelationships among contributing factors; and (4) propose a conceptual prototype of an AI-driven, Ultra-advanced Collision Avoidance System (AUCAS-L3) targeting HVI-induced driver distraction under automated driving without eye-tracking and video-recording. Fixation and pupil dilation data from the eye tracking device are used to model driver distraction, focusing on visual and cognitive distraction, respectively. In order to validate the proposed methods for measuring and modeling driver distraction, a data collection was conducted by inviting drivers to try out automated driving under Level 3 automation on a simulator. Each driver went through a jaywalker scenario twice, receiving a takeover request under two types of HVI, namely “visual only” and “visual and audible”. Each driver was required to wear an eye-tracker so that the fixation and pupil dilation data could be collected when driving, along with driving performance data being recorded by the simulator. In addition, drivers’ demographical information was collected by a pre-experiment survey. As a result, the magnitude of visual and cognitive distraction was quantified, exploring the dynamic changes over time. Drivers are more concentrated and maintain a higher level of takeover readiness under the “visual and audible” warning, compared to “visual only” warning. The change of visual distraction was mathematically formulated as a function of time. In addition, the change of visual distraction magnitude over time is explained from the driving psychology perspective. Moreover, the visual distraction was also measured by direction in this research, and hotspots of visual distraction were identified with regard to driving safety. When discussing the cognitive distraction magnitude, the driver’s age was identified as a contributing factor. HVI warning type contributes to the significant difference in cognitive distraction acceleration rate. After drivers reach the maximum visual distraction, cognitive distraction tends to increase continuously. Also, this research contributes to quantitatively revealing how visual and cognitive distraction impacts the collision hazards, respectively. Moreover, this research contributes to the literature by developing deep learning-based models in predicting a driver’s visual and cognitive distraction intensity, focusing on demographics, HVI warning types, and driving performance. As a solution to safety issues caused by driver distraction, the AUCAS-L3 has been proposed. The AUCAS-L3 is validated with high accuracies in predicting (a) whether a driver is distracted and does not perform takeover actions and (b) whether crashes happen or not if taken over. After predicting the presence of driver distraction or a crash, AUCAS-L3 automatically applies the brake pedal for drivers as effective and efficient protection to driver distraction under automated driving. And finally, a conceptual prototype in predicting AV distraction and traffic conflict was proposed, which can predict the collision hazards in advance of 0.82 seconds on average.

Download

Share

COinS