Predicting cruising speed through data-driven driver modeling

We present a data-driven method for predicting driver behavior of sufficiently low complexity to be implemented in an automotive context. In this work, we develop a method to predict the driver's intended cruising speed as they launch from a stopped position. Our goal is to make this prediction in spite of highly modal driving by the driver (i.e. they drive in either an aggressive or relaxed manner). To reduce complexity and improve prediction, we do not try to calculate the hidden variables causing the modal driving or try to predict the vehicle's entire trajectory through filtering. We instead formulate a supervised learning problem to estimate the cruise speed directly. First we segment the trajectories into launch, cruise, and deceleration behavioral segments based on vehicle state and environment. Within each of these behavioral segments, we extract a low dimensional feature set which can be used to learn a model for predicting cruise speed under modal driving. In particular, a dynamical model is fit to the launch sequence data and then the coefficients of the model are used as regressors for a Nadaraya-Watson estimator. The method is implemented real-time in a vehicle, and results show that for a single road type, prediction error is significantly lower than other standard prediction methods. A key point of this paper is that our simpler prediction technique can yield good prediction results over long time scales with low complexity by predicting goal states directly rather than predicting the evolution of the vehicle state in time.