Driver state estimation for prediction of vehicle states within control systems

In this paper we present an online learning approach to predict driver behavior and resulting vehicle states. The driver is represented by driver states and a control function. Kernel Density Estimation is used for online estimation of current driver states. Data sampling methods are introduced to observe the virtual driver states. The driver states are used as input for the control function to predict resulting vehicle states. To consider environmental influence on driver behavior a context-separated learning approach is presented. The system is tested with real drive test data from different drivers on a specified test route. Different settings regarding learning speed and type of context-separation are investigated. Results show, that consideration of environmental influences on the driver states lead to a better identification of the current behavior but prediction on a longer time horizon does not necessarily improve correspondingly.