Matching with Feature Segments of Stereo Images by IA

Abstract

Matching of stereo images is a fundamental task for 3D recovery. We have matched with feature points of a direct image and its mirror image by using GA and the information of apexes connections. However, when the feature points become large, miss matching were occurred because of the graph imbedding problem and the ambiguities of matching with occluded points. We introduce an immune algorithm (IA) to cope with these multi-peaks searching.
In this paper, we propose a matching algorithm of feature points/segments by using cut set graph and IA. First, the stereo images are specified to a criterion and a reference image, respectively. Both the feature points/segments of the images are classified to 3 groups; cut set, inner and outside group. An evaluation of reference image is defined how many the remark points/segments are included in each groups correctly correspond to these of the criterion groups, i. e, the evaluation is the degree of the fitness. The candidates of matching obtained from IA are selected by checking both the graphs connection conditions. We show that the peaks are effectively detected and matched correctly.