Modular stereo vision - theory, implementation, and application
2017-02-17T03:07:16Z (GMT) by
Just like the biological eye, human visual depth perception is extremely complex. Part of this complexity is that it is multi-modal. One of the primary modes is that of stereo vision. Computational stereo vision aims to replicate this mode of depth perception. The research in this thesis is focused on improving computational stereo vision. The stereo vision problem is primarily a search/matching problem. Many stereo vision algorithms have been proposed to solve the stereo matching problem. Several paradigms are available to make sense of and differentiate between these algorithms. This thesis considers the relative merits and utility of the current paradigms. The component analysis paradigm is selected as being best suited to both classify and improve on current stereo vision algorithms. The Modular Stereo Vision Model (MSVM) is presented. This model is based on the component analysis paradigm. The MSVM provides a single set of classification rules which enables accurate segmentation of current and future stereo vision algorithms. A programmatic implementation of the MSVM, called Anystereo, is also presented. Anystereo is a complete system which allows arbitrary stereo vision methods to be combined to form an entire stereo vision algorithm. The Anystereo system is used to implement three novel stereo vision methods. The Intrinsic Images method deals with illumination variation between the images of the input stereo pair. This method is based on identifying and isolating the intrinsic components of the input images which do not depend on illumination. The results of applying this method in the presence of illumination variation are presented. The Quartile Matching method deals with exposure variation between the images of the input stereo pair. This method is based on the observation that the relative ordering of intensities does not change when exposure varies. A non-parametric ranking is used to measure matching cost between the input images. The results of applying this method in the presence of exposure variation are presented. The Connected Component Aggregation method deals with the boundary problem. The boundary problem arises at disparity boundaries when one surface is more highly textured than the other surface. This method aggregates over disparity-slices rather than over the pixels of the input images to prevent aggregation from crossing disparity boundaries. The results of applying this method are presented. The work presented here is targeted at improving the stereo vision field of research. This is accomplished practically through various stereo vision methods and philosophically by the introduction of the MSVM and the Anystereo system.