Assessing Driver Engagement Through Machine Learning Classification of Physiological Measures

The assessment to determine if drivers are engaged and supervising autonomous and semi-autonomous vehicles (AV) is becoming an increasingly important task. AVs are becoming more prevalent on the road and adequate driver engagement is a must for safe operation. The full adoption of level 5 vehicles will likely take many years as they do not currently exist for consumer purchase and before this happens the roads will see a mix of level 0-4 vehicles. During this time humans will still be responsible for taking control of the vehicle if a hazardous scenario occurs and the AV does not know how to maneuver around. To have a safe handover from AV to human during these situations the driver must maintain a level of engagement even while the AV is driving. To do this physiological sensors can be used to measure signals such as heart rate and respiration rate, which are known indicators of a driver’s engagement. This paper exposes drivers to non-surprise and surprise driving scenarios to assess if attentive drivers can be identified from physiological changes for manual and AV driving. Machine learning (ML) is used to understand the patterns of physiological signals and classify when a driver is engaged during a surprise scenario. Finally, the ML models show a successful ability to classify engaged versus non-engaged drivers with a 73.3% accuracy in manual driving scenarios, 86.7% accuracy in AV, and 70.0% accuracy when the data from both driving scenarios are combined showing the model’s ability to generalize.