A model-free voting approach to cue integration

Vision systems, such as "seeing" robots, should be able tooperate robustly in generic environments. In this thesis, weinvestigate certain aspects of how these demands of robusmessof a systems approach to vision could be met. Firstly, we suggest that robustness can be improved byfusing the variety of infor mation offered by the environment,and, therefore, we investigate the effectiveness of using thecoincidence of multiple cues. Secondly, we are concerned aboutthe use of coarse algorithms. Even though the environmentprovides much information, it is neither necessary nor possibleto extract all information available. Therefore, we will showthat coarse algorithms will suffice for certain problems. Toinvestigate the effectiveness of using the coincidence ofmultiple cues, we perform a series of experiments on detectingplanar surfaces in binocular images. These experiments arebased on two schemes of a somewhat different character. The first one is a hyporhesls-and-test scheme that incorporates the cues ina certaln order and hence, by design, imposes a ranking ofthem. The general idea is to use arbitrary cues exploitinglocal image data to get an idea about whether the model (aplanar surface) is seen in the image and at which location itis found. If one or more cues strongly indicate a certaininstance of a model, then this observation serves as ahypothesis to be tested by other cues to support or reject thishypothesis. In comparison to the cues used for hypothesisgeneration, those used for hypothesis testing should be morereliable and can also have a higher computational complexitysince they are only employed when needed.The general idea of the second scheme is to first use asimple, and quick cue exploiting local image data toget anidea of where in the image the model (a planar surface) couldbe found. After this initial localization step, all cues thatcan be computed are gathered and allowed to vore for the occurrence of the model in the hypothesizedregion. The initialization of this approach is a hypothesisforming step, similar to that of the hypothesis-and-testapproach This step though, is much weaker because it onlyindicates a region in the images where to look. The approachallows direct fusion of incommensurable cues, such as intensityand surface orientation. Generally, it can be regarded as aless restrictive approach than the hypothesis-and-test approach. We propose that coarse algorithms may be motivated from arobustness and flexibtl hy point of view. Our exp