Modeling and Prediction of Human Driver Behavior

Knowledge of the current and future driving context could facilitate the interaction between human driver and advanced driver assistance systems. A driver's intended actions (the future context) can be inferred from a number of sources, including the driver's current control actions, their visual scanning behavior, and the traffic environment surrounding them. In an approach similar to hidden Markov models, the intended actions (e.g., to turn or change lanes) are modeled as a sequence of internal mental states, each with a characteristic pattern of behavior and environmental state. By observing the temporal patterns of these features, it is possible to determine which action the drivers are beginning or intending to execute. This approach has been successfully demonstrated in a variety of simulated driving conditions for a wide range of driver actions including emergency maneuvers. In these studies, only the control actions of the driver (i.e., steering and acceleration actions) were used to infer the driver's state. We are presently exploring the use of the driver's visual scanning behavior as another source of information about the driver's state. Visual scanning behavior offers the additional advantage of prediction of driver actions since scanning generally takes place in areas ahead of the current car position.