Studies in Robotic Vision, Optical Illusions and Nonlinear Diffusion Filtering

Abstract: This thesis is divided into three parts, which all deal with computational analysis and processing of images. However, the settings are quite diverse.They range from robotic camera sensors, over human perception, to physical measurement setups. In spite of this diversity, common theoretical ideas and computational aspects are applied which ties the three different parts closer together. The first part of the thesis deals with calibration methods for robotic vision. Both the estimation of the intrinsic parameters of the applied camera model, intrinsic camera calibration, and the estimation of the orientation and position of the camera in relation to the end-effector of the robot, hand-eye calibration, are discussed. Two different methods are presented. The first one explores the constraints that arise when calibrating a single camera or a stereo head configuration using a planar calibration object, while performing translational or general motions. The other one uses estimations of the spatial and temporal intensity derivatives in an image sequence for direct computation of the unknown parameters. The second part of the thesis discusses a new framework for explaining a number of geometrical optical illusions. It is proposed that noise, that enters into the visual process at different stages, causes the estimation of different features in the observed image to be biased. Different types of error models are discussed and illusions that are best explained by each particular model are presented. The discussion is not restricted to the human visual system and highlights the importance of analyzing the influence of noise and uncertainty in any visual process. The third and final part of the thesis propose the use of nonlinear diffusion filtering to process images obtained by planar laser-induced fluorescence (PLIF) spectroscopy. In particular, the images in the present application are PLIF images of turbulent flames in combustion processes. Solving a nonlinear diffusion equation using an image of this type as initial value, makes succeeding extraction of interesting quantities, such as the length of the flame boundary, an easy task. An analysis of the properties of nonlinear diffusion filtering in general, and for the present application in particular, is given.