Abstract
In order to develop intelligent active safety systems for vehicles, accurate and precise models of the driver must be developed. This is because while the driver has full control of the vehicle, her likely actions are highly unpredictable. In this chapter, we present an individualized, probabilistic model that predicts driver behavior based on the outside environment, the state of the driver, and previously observed trajectories. The overall goal of this driver model is to create an individualized prediction scheme over a long time horizon in various situations by using empirical driver data to create a set of potential future trajectories, which is inspired by reachable set techniques (a standard safety metric used in control). The method integrates clustering algorithms to identify scenarios, parse through sensor data, and collect the associated driver behavior in a way that is readily used in control applications. This model can then be used to determine when a safe controller should intervene to keep the vehicle in a safe region of the current environment in a probabilistic manner. Experimental results from human-in-the-loop testbeds show that this method has extremely high accuracy, in both highway and city intersection scenarios. Data analysis of the model performed to examine the reliability of the model, characterize human behavior, and quantify the dissimilarities of driver behavior in different scenarios. By using this flexible methodology, a precise and accurate individualized driver model can be developed for use in semiautonomous frameworks.
