Predicting Environmental Demand and Secondary Task Engagement using Vehicle Kinematics from Naturalistic Driving Data

In this paper, we focus on exploiting the vehicle kinematic signals from a naturalistic driving dataset to estimate motor control difficulty associated with the driving environment (i.e., curvature and poor surface condition), and detect whether the driver was engaged in a secondary task or not. Advanced driver assistance systems can exploit such driver behavior models to better support the driving task and improve safety. Hidden Markov models were built from sequential data; lateral (x-axis) and longitudinal (y-axis) acceleration were used to classify motor control difficulty (lower vs. higher), whereas GPS speed and steering wheel position were used to classify secondary task engagement (yes vs. no). The resulting accuracy for lower motor control difficulty classification was 72.03%, whereas 71.27% was achieved for higher motor control difficulty. Cases of engagement in secondary task vs. not were classified with 84.4% and 74.0% accuracy, respectively.