Classifying Cognitive Load for a Proactive In-car Voice Assistant

Using Personal Assistants (PAs) via voice becomes increasingly popular and available in multiple environments, thus we aim to provide proactive PA suggestions to car drivers via speech. However, these suggestions should not be obtrusive or cognitively overloading the driver during the interaction, to regard road safety. Consequently, we need to model proactive dialogs related to drivers' cognitive load. To reach this goal, we classify different levels of drivers' cognitive load. We take a multi-step approach: First, we collect real-time CAN data from a Wizard of Oz driving simulator study (i.e., brake pedal velocity, steering wheel angle, steering wheel acceleration and speed). Then we apply unsupervised clustering to identify different cognitive load levels. Four clusters are obtained and labeled accordingly: low, medium, medium-high and high cognitive load. After conducting feature generation, feature selection, and resampling, we apply different classification algorithms. By combining SVMs and SMOTE we achieve an accuracy of 96.97%.