Predictive Modeling of Driver Risk Perception Utilizing Driving Simulator: A Deep Learning Approach

Abstract: The ability and performance of drivers can vary greatly between drivers and even for the same driver in different driving conditions. These variations are influenced by elements such as the driver's mental workload (MW), the difficulty of the task (TD), perceived risk (PR), and other factors like motivation, exhaustion, drowsiness, drug use, distractions, and stress. Driver’s MW combined with TD can influence PR that might not be similar for two persons in the same driving condition. As such, people with high-risk perception are more likely to feel less risky in similar situations and have more control over the situation, compared to other drivers. The goal of this study is to model the real-time PR of a driver by mapping the driver’s traits, vehicle trajectories, as well as physiological measures. Modeling PR can be useful in the development of adaptive systems, effective training programs, and policy decisions related to road design, signage, and designing vehicles and road systems more compatible with human capabilities. The braking behavior serves as a reliable continuous measure for assessing the level of PR, as braking is used to compensate for the excess MW. The brake pedal force indicates the amount of risk that the driver intends to reduce by applying pressure to the brake pedal which is contingent upon their current level of skill and the level of workload demand or task difficulty in that particular situation. Risk Perception (RP) is a time sequence phenomenon that can be regarded as the output of human cognition that results in response to the surrounding environment and situation. As deep learning techniques can mimic human behavior and capture the patterns from time-series data, a novel deep-learning algorithm called Long Short-term Memory (LSTM) was used to build the RP model. The data for this research were collected using the KU driving simulator, a fixed-based simulator in an Acura MDX chassis (half cab) which also includes an eye-tracker. The driving scenario incorporated six tasks with varying levels of difficulty due to varying traffic density, lane changing/deviation activity, heavy vehicle presence, and the number of lanes. The brake pedal force was used as an indicator of PR that was converted into scores ranging from 0 to 50. The model used the current and previous driving states such as vehicular trajectories as well as eye-tracking parameters (eyelid opening, pupil diameter) and driver’s demographic information (age, gender, and driving experience) as inputs and predicted the PR for the 1 step (0.6 sec) into the future. The model produced reliable prediction accuracy with an RMSE value of 0.02. Using this RP model, a driver can perceive the incoming risk and act accordingly on time. Additionally, the RP of different driver groups in different driving conditions can be explored using this model. Moreover, the model will open a new door to personalize future vehicles according to one’s driving behavior and personal traits.