Risk-Aversive Behavior Planning under Multiple Situations with Uncertainty

摘要

This paper addresses the problem of future behavior evaluation and planning for upcoming ADAS, especially for inner city traffic scenarios. Situations in inner city traffic scenarios are generally highly complex and of high uncertainty. The behavior in such complex scenarios differs strongly depending on the actually occurring situation. In general the current situation can only be determined with high uncertainty based on current and past sensory measurements of the ego entity and the other involved entities. Additionally a situation can change very quickly, e.g. if a traffic participant suddenly changes its behavior. Here we propose an approach how to plan safe, but still efficient future behavior under consideration of multiple possible situations with different occurrence probabilities. For each situation we predict prototypical future trajectories of all involved entities using a highly general, interaction aware model Foresighted Driver Model (FDM). Then, based on a continuous, probabilistic model for future risk, we build so-called predictive risk maps, one for each possible situation, and plan the own behavior while minimizing overall risk and utility. We show that our approach generates efficient behavior for situations with high probability, while generating a "plan b" to safely deal with improbable but risky situations.