Driver workload in an autonomous vehicle 2019-01-0872
It is posited that the occupant of an autonomous vehicle (AV) is neither workload nor stress free. Even though an autonomous vehicle may operate under rule-based sensors, actuators and algorithms, the human occupant’s vestibular, kinesthetic and ocular control system, whether age and/or experience related, may not be in synch with the vehicle. As a result, there may likely be discomfort, dissatisfaction, anxiety and unease for the human. Earlier research showed that an occupant’s perceived workload could be used as a surrogate to reflect this discomfort. This study further refines the determination of perceived workload and suggests how OEMs could use perceived workload in AV design.
Perceived workload was determined based on occupant physiological measures. Because of great variation in individual personalities, age, driving experiences, gender, etc., a generic model applicable to all could not be developed. Rather, individual workload models that used physiological and vehicle measures were developed. Unlike the methods of workload estimation where a single, or a few signals are used, such as electroencephalography (EEG), electrocardiography (ECG), we used several signals to generate a more promising performance; an end-to-end deep neural architecture that makes workload estimation using a combination of physiological and vehicle signals. The architecture receives multiple heterogeneous temporal signals simultaneously, learns features and makes classifications. All data collected for training and testing are from real driving scenarios including local urban and highway, so the concern about the discrepancy in data acquisition between simulator and real driving is avoided. Data from ten participants ranged from new drivers to those who had been driving for decades were collected and analyzed in this study. The experimental results indicate that the proposed driver workload estimation model is capable of learning well from the combined temporal physiological and vehicle signals and obtains good performance on workload estimation.
Yi Murphey, Dev S. Kochhar, Yongquan Xie
Univ of Michigan-Dearborn, Ford Motor Company, University of Michigan-Dearborn