Determining Potential Obstacles in Unobservable Areas Based on Current and Past Perception

Automated vehicles perceive only a small part of their environment. Especially unobservable vehicles pose a significant risk. To achieve safe but also comfortable behavior, potential, unobservable vehicles must be considered in behavior planning. Conventional methods use solely the current observation of the environment to determine potential obstacles. Past observations are rarely considered, although these may contain helpful information to rule out potential obstacle positions. This paper presents a novel algorithm that uses past observations besides the current observation to determine possible obstacle states. By means of a particle filter, we iteratively predict and filter feasible states of a potential obstacle. This results in a probability distribution for the position and velocity of an unobservable obstacle. We furthermore present a concept for the interface between our method and a basic behavior planning algorithm. The real-time capable method is tested on both simulated and real-world data. By comparing the algorithm to a baseline algorithm which uses only the current observation, we show that our algorithm prevents overly cautious assumptions about a potential obstacle’s state in certain situations. As a result, a more comfortable driving behavior can be achieved.