Prediction of Object Position based on Probabilistic Qualitative Spatial Relations

Abstract

Due to recent and extensive advancements in the robotic and artificial intelligence fields, intelligent systems can be found, with increasing frequency, in many areas of daily life. From industrial and surgical purposes to space robots, such complex systems are present. However, as demands for robotics systems increase, sophisticated algorithms for use in robotic areas such as perception, navigation, or manipulation are required. Although some algorithms for such purposes exist, there are still open questions and challenges that must be addressed. Although robots are primarily used in the manufacturing industry, which has since been revolutionized by their precision and speed, there is a growing trend towards using service and personal robotics applications. The latter in particular must interact with humans naturally and effectively manage their environments, such as offices and homes. In contrast to the systems used in an industrial context, systems such as personal robots do not act in a predefined and fixed environment. Rather, these intelligent systems need an intrinsic comprehension of human environments to be able to support people in their daily life and manage common tasks such as preparing a breakfast table or cleaning a room. Crucially, these new robot systems require an entirely new level of capabilities to act in dynamic human environments. This thesis addresses how qualitative spatial relations can be used to find an objecta s most probable location and thus guide the search for a sought object. Because current approaches focus mainly on crisp, two-dimensional relations, which are not directly suitable for use in three-dimensional real-world applications, a formalism for a new type of spatial relations is proposed in this work. This theoretical approach is then applied on real-world data to evaluate its applicability for robotics purposes. The resulting validation of the approach demonstrates that the developed method performs well and can be used to enhance search for objects.