Traffic Situation Assessment by Recognizing Interrelated Road Users

With the trend to highly automated driving, future driver assistance systems are required to correctly assess even complex traffic situations and to predict their progress. As soon as other road users are present the number of possible situations becomes infinite, rendering their assessment based on learned situation types impossible. In this paper we propose to break the situation down into sets of interrelated entities by estimating for each road user the entities that affect its behavior most. The decomposition offers numerous advantages: Attention can be focused on relevant entities only and predictions can be performed with a smaller set of considered entities. As the high variability among situations requires a large amount of data for learning and testing, we implemented a simulation environment that gives access to the causes for the behavior of each road user. In a simulated intersection scenario we show that we can reliably infer the affecting entities for each road user only utilizing features that can be obtained by common sensors.