Feature Relevance Estimation for Learning Pedestrian Behavior at Crosswalks

For future automated driving functions it is necessary to be able to reason about the typical behavior, intentions and future movements of vulnerable road users in urban traffic scenarios. It is crucial to have this information as early as possible, given the typical reaction time of human drivers. Since this is a highly complex problem, it needs to be addressed in small portions. In this paper we will focus on the behavior of pedestrians at crosswalks. We use a database of real pedestrian trajectories to learn a model which is able to predict if a pedestrian will cross the street. Therefore, we first introduce a large set of possible features that could be suitable to describe the behavior. Afterwards, we perform relevance determination to identify those features that are necessary to reach the best possible generalisation performance. We provide experimental results on data collected at a pedestrian crossing in a city in southern Germany. Our results shows, that a very sparse set of features, which depends only on the pedestrians' trajectory, gives the best result.