Aspects of 3D Perception, Abstraction, and Interpretation in Autonomous Mobile Robotics

Abstract

This thesis presents novel insights and solutions to the problem of perceiving, abstracting, and interpreting the 3D structure of human environments with autonomous mobile robots. It focuses on the goal of retrieving a meaningful representation of objects from 3D laser range data for purposes of object recognition. As such, it represents an important contribution to the development of cognitive technical systems, which depend on intelligent perception as a core capability. A chain of abstraction is developed throughout the chapters, that leads from raw point cloud data to recognized objects. Following a thorough review of state-of-the-art sensing approaches, new studies and methods for feature computation, surface-based point cloud segmentation and Bayesian part-based object recognition are presented. The results show that with the framework a large number of structured objects can be learned and robustly recognized from few examples.