Scene interpretationand localization has been identified as one of the most fundamental and challenging tasks already since Shakey. In recent years the problem has often been addressed independently with varying foci within AI, vision, and robotics. But new ambitious (benchmark) tasks like ambient intelligence and wearable computing had shown the necessity of integration of formerly independent solutions. This integration process requires solutions to some new inherent difficulties: E.g., localization is one of the most fundamental tasks for AAL that also constitutes for the map generation of the surrounding environment (e.g., by SLAM approaches). The currently successful methods (i.e., Monte-Carlo, (extended) Kalman-Filter) handle sensor noise based on a set of probabilistic methods which directly result in a probabilistic representations of the environment. In contrast most classic knowledge representation techniques are based on strict non-probabilistic representations. Generally, the resulting problem can be addressed in at least three different ways: (1) Minimizing uncertainty and (trying to) generate a strictly declarative representation with full inference power. (2) Handling uncertainty by the generation of a hybrid representation which incorporates probabilistic and declarative representations at the expense of less powerful declarative inferences. (3) Or directly representing uncertainty by the use of a strictly probabilistic representation. Each approach imposes strengths as well as (strict) limitations on the inferences available. In our presentation we introduce a qualitative approach to navigation, based on ordering information which helps to overcome these limitations. Since the complete localization task is based on purely qualitative representations a probabilistic representation is not necessary. Furthermore we show that qualitative localization based on ordering information provides strong inherent constraints which support robust localization even in the face of uncertainty based on an extremely limited set of information.

-> Download presentation