Spatial Representation and Reasoning for Robot Mapping

This thesis addresses spatial representations and reasoning techniques for mobile robot mapping. It provides an analysis of fundamental representations and processes involved. A spatial representation based on shape information is proposed and appropriate shape analysis techniques are developed. Robot mapping’s core problem of determining correspondences between observation and map is tackled on the basis of shape similarity. An improved matching technique is described by a generalized mathematical formulation. Specifically, it addresses the matching of configurations of extended geometric primitives’ configurations. Robot mapping describes the process of a robot autonomously acquiring an internal spatial representation of its environment. This internal representation, commonly termed “map”, provides the basis for planning future actions. A reliable map is essential for intelligent navigation. Henceforth, autonomous map acquisition is one of the most fundamental tasks for autonomous robots. Unfortunately, it is among the most challenging tasks as well. To build a map, robots rely on observations, typically obtained from their own sensors, but sometimes information can also be obtained by communication. The map is constructed by integrating multiple views on the same spatial environment into a coherent whole. This requires the correlation of observations, i.e. to determine which observations refer to the same physical entity—this is the objective addressed by the so-called correspondence problem. Additionally, multiple, corresponding observations need to be integrated into a single model; this task is termed the merging problem. There are several aspects contributing to the difficulty of robot mapping. Among them, first of all, is the development of a robust and efficient solution to the correspondence problem. Another aspect is the necessity to handle uncertain information; virtually all information available to the robot must be considered uncertain. For example, sensor readings suffer from noise and undetermined failure which results in uncertain information by interpretation. Besides the effects of uncertainty to observations, the environment may simply change between observations, complicating recognition. Despite the high complexity faced in the robot mapping problem, a real-time solution is indispensable in many applications. My work is dedicated to improving spatial representation and reasoning techniques underlying the robot mapping task. I argue for utilizing a shape representation originating in the field of object recognition; a strong, yet underexploited connection between the research fields of shape recognition and robot mapping is explored. Distinctive shape similarity information facilitates an efficient and robust approach to the correspondence problem. The goal is