Description Logic for Vision-Based Intersection Understanding

Abstract — Road recognition from video sequences has been solved robustly only for small, often simplified subsets of possible road configurations. A massive augmentation of the amount of prior knowledge may pave the way towards a generation of estimators of more general applicability. This contribution introduces Description Logic extended by rules as a promising knowledge representation formalism for scene understanding. A Description Logic knowledge base for arbitrary road and intersection geometries and configurations is set up. Logically stated geometric constraints and road building regulations constrain the hypothesis space. Sensor data from an in-vehicle vision sensor and from a digital map provide evidence for a particular intersection. Partial observability and different abstraction layers of the input data are naturally handled. Deductive inference services – namely classification, entail-ment, satisfiability and consistency – are then used to narrow down the intersection hypothesis space based on the evidence and the background knowledge, and to retrieve intersection information relevant to a user, i.e. a human or a driver assistance system. The paper concludes with an outlook towards non-deductive inference, namely model construction, and pro-babilistic inference. I