Deciding what to inspect first: Incremental situation assessment based on information gain

In order to offer even more sophisticated functionality, future driver assistance systems need the ability to robustly recognize and understand driving situations. Especially in inner-city scenarios the high complexity and variability of situations encountered make their assessment a challenging task. We propose to tackle these challenges by decomposing situations into smaller, more manageable parts. We define such a part as a set consisting of a road user and all entities (e.g. cars, traffic lights) currently affecting its behavior. Though the decomposition alleviates the assessment already, for higher numbers of present entities the recognition of interrelated entities is still computationally expensive if performed in a brute-force fashion. Therefore we employ sensitivity analysis on Bayesian Networks for sensibly controlling the recognition process on the basis of information gain. This leads to an active measurement process in which a situation is perceived incrementally, concentrating first on the most meaningful sensor measurements. The proposed method is evaluated on a simulated inner-city scenario where it reliably recognizes the affecting entities of each road user. We show that a recognition process based on information gain can save more than 50% of measurements without significantly impairing the recognition rate.