The introduction of automated driving motivates the need for driver state detection, prediction and monitoring. The first introduced systems for automation (SAE level 2 or 3) rely on the ability of the driver to resume the driving task, both manually (steering control) and visually. The automation system monitors the presence of the driver’s hand(s) on the steering wheel and some systems monitor the driver’s visual attention as well. This is necessary to ensure that the driver monitors the automated drive and responds manually and/or visually to HMI cues from the automation system.
Considering higher levels of automated driving (SAE level 3 or 4), several challenges and opportunities arise. First, as driver confidence in the performance of automation grows, they will switch attention to secondary tasks increasing their cognitive workload. The driver may be busy doing other tasks and a system which estimates cognitive workload can indicate that a longer take-over period is necessary, or, in the extreme, a human driver take-over may not be safe or even feasible. Secondly, the driver may be visually, manually or mentally overloaded (or a combination of these) during manual control and the automation system might encourage or intervene automation modes to enhance safety. These two use cases can be improved using accurate real-time prediction of the driver’s mental workload using one or more in-vehicle sensors.
In this paper a robust method for estimating driver workload based on real in-vehicle sensors is presented. Sensors in the seatbelt and at the steering wheel rim derive heart rate metrics which are used to estimate cognitive workload. Furthermore, an analysis is conducted to determine if additional metrics derived from vehicle dynamics data have an impact on the calculation accuracy. Comparing individual-based driver classification approaches versus a generalized driver algorithm is also part of this investigation. A driving simulator study with n-back task induced workload is used to validate the driver cognitive workload estimation method accuracy.