Description Logic for Scene Understanding

Understanding a natural scene on the basis of external sensors is a task yet to be solved by computer algorithms. Solving it for non-microworlds requires an appropriate formalism for representing and reasoning with sensor data and domain knowledge. This contribution investigates the suitability of the family of Description Logics (DL) for this task. DL are an offspring of First Order Logic, inheriting its well-defined, declarative semantics while offering improved computational tractability and object-oriented knowledge engineering. The first part of this contribution elaborates on principled approaches on using DL for representing scene information and for solving scene understanding tasks. The second part describes an extensive case study in the application domain of road intersections, introducing RONNY, the ROad Network oNtologY. RONNY is evaluated on a set of complex intersection scenes, using as input data an experimental vehicle's camera platform, digital map and GPS sensor. The experimental results support the claim that a logic-enhanced system can improve state-of-the-art quantitative Computer Vision, and enables tackling tasks that are beyond the current scope.