Clustering of drivers’ state before takeover situations based on physiological features using unsupervised machine learning

Conditionally automated cars share the driving task with the driver. When the control switches from one to another, accidents can occur, especially when the car emits a takeover request (TOR) to warn the driver that they must take the control back immediately. The driver’s physiological state prior to the TOR may impact takeover performance and as such was extensively studied experimentally. However, little was done about using Machine Learning (ML) to cluster natural states of the driver. In this study, four unsupervised ML algorithms were trained and optimized using a dataset collected in a driving simulator. Their performances for generating clusters of physiological states prior to takeover were compared. Some algorithms provide interesting insights regarding the number of clusters, but most of the results were not statistically significant. As such, we advise researchers to focus on supervised ML using ground truth labels after experimental manipulation of drivers’ states