Collision Avoidance for Autonomous Cars Based on Human Intention

Abstract

This thesis considers a problem of controlling an autonomous car cooperating with human-drivencars. Proposed proactive collision avoidance system incorporates human-driver�s intentions transfered viaVehicle-to-Vehicle (V2V) communication. This system utilizes multi-step Vector Gaussian Processes (VGP)and Stochastic Transition models in order to learn the resulting transitions for each given intention and updatethem on-line from the real driving scenarios to provide an adaptive intention-based trajectory predictionfor any kind of driving manner and road/weather condition. Such an approach allows us to use stochasticbehavioral model for a numerical evaluation of the risk of collision and pose the question of control of thevehicle as an optimization problem. This formulation makes it possible to utilize various existing optimizationtechniques what is proven by the use of both a single-step cost function minimization and sequentialdecision making algorithm based on Markov Decision Process (MDP). The effectiveness of this concept issupported by a variety of simulations utilizing the real human driving in various scenarios considering bothan intersection and highway scenarios in specially developed Matlab driving simulation and highly realisticthird-party developed car simulator Carnetsoft.