Integrative Relational Spatial Representation and Reasoning

Abstract

Qualitative Spatial Representation and Reasoning (QSRR) provides a qualitative and relational understanding of our spatial world, where each scenario is described by the set of qualitative spatial relations between existing objects. Spatial intelligence achieved by QSRR consists of at least two parts: the ability to represent spatial aspects that are indispensable for modeling a certain problem and the power of deriving new knowledge for problem solving based on information at hand. Qualitative spatial representation models consist of qualitative spatial relations in various forms. For example, qualitative topological relations between regions and qualitative direction relations between points. Qualitative spatial reasoning techniques enable converse reasoning, compositional reasoning, consistency checking, and neighborhood-based reasoning. The main advantage of QSRR is that it is close to the way humans de- scribe spatial scenarios and reason about spatial problems. QSRR has been applied in domains like navigation and action description.

Previous research on QSRR pays more attention on theoretical analysis of rep- resentation models and reasoning techniques rather than practical issues when applying QSRR in solving real-world problems. To achieve better practical usefulness, qualitative spatial relations should be internally integrated with each other according to their implicit connections; and externally integrated with techniques from other relevant fields. In this thesis, we focus on solutions on four specific research problems to enlighten future research on achieving better practical usefulness of QSRR by integration. First, integrating qualitative direction relations that are described in the same or different frames of reference. We discuss and formally define three different categories of frames of reference; propose a set of rules that enables across-frames-of-reference converse reasoning and compositional reasoning; and define the constraint satisfaction problem in the context of direction relations across different frames of reference. Second, integrating spatial and non-spatial relations between landmarks to achieve human-like localization and navigation. We define qualitative maps to store qualitative relations between places; propose an interactive localization algorithm that is capable of locating agents based on eliminating unsatisfying landmarks; and design a reliable navigation system which takes number of landmarks and qualitative distance information into consideration. Third, integrating structured symbolic spatial representation with unstructured numeric learning methods. We propose an innovative neuro-symbolic approach that recognizes actions from videos with explanations. These explanations are in the form of qualitative spatial object relation chains and thus human-understandable. Forth, integrating various spatial representation models to achieve an efficient model selection strategy. We explore solutions for this problem and propose three different representation determination strategies. Show more