A model-free voting approach for integrating multiple cues

Computer vision systems, such as “seeing” robots, aimed at functioning robustly in a natural environment rich on information benefit from relying on multiple cues. Then the problem of integrating these become central. Existing approaches to cue integration have typically been based on physical and mathematical models for each cue and used estimation and optimization methods to fuse the parameterizations of these models. In this paper we consider an approach for fusion that does not rely on the underlying models for each cue. It is based on a simple binary voting scheme. A particular feature of such a scheme is that also incommensurable cues, such as intensity and surface orientation, can be fused in a direct way. Other features are that uncertainties and the normalization of them is avoided. Instead, consensus of several cues is considered as non-accidental and used as support for hypotheses of whatever structure is sought for. It is shown that only a small set of cues need to agree to obtain a reliable output. We apply the proposed technique to finding instances of planar surfaces in binocular images, without resorting to scene reconstruction or segmentation. The results are of course not comparable to the best results that can be obtained by complete scene reconstruction. However, they provide the most obvious instances of planes also with rather crude assumptions and coarse algorithms. Even though the precise extent of the planar patches is not derived good overall hypotheses are obtained. Our work applies voting schemes beyond earlier attempts, and also approaches the cue integration problem in a novel manner. Although further research is needed to establish the full applicability of our technique our results so far seem quite useful.