Prediction of Driving Behavior through Probabilistic Inference

Abstract: Driving assistance systems are essential technologies to avoid traffic accidents, reduce traffic jams, and solve environmental problems. Not only observable behavioral data, but also unobservable inferred values should be considered to realize advanced driving assistance systems that are adaptable to individual drivers and situations. For this purpose, Bayesian networks, which are the most consistent inference approach, have been applied for estimation of unobservable physical values and internal states introduced for convenience’s sake. Nevertheless, only a few reports have addressed prediction of future states of driving behavior. This paper proposes predicting driving behavior in the near future through a simple dynamic Bayesian network, which is a hidden Markov model or a switching linear dynamic system. The proposed predictors were examined with real data. We focused on prediction of the future stop probability at an intersection because it is one of the most important maneuvers for safety to avoid collision with other traffic elements (i.e. other vehicles and pedestrians) at an intersection. Both the HMM and the switching linear dynamic system worked well as stop probability predictors. The HMM represented the temporal structure of human driving behavior.