Matching of featured objects using relational tables from stereo images

A method for three-dimensional scene matching is presented in this paper. Both the geometric and structural information of the segmented features in two images are used for three-dimensional scene matching. The segmented features such as regions, edge segments, and vertices are initially labelled by using a symbol set. Then the structural relationships among these labels in each image are tabulated in a relational table. The consistent labels between two relational tables associated with two given images are searched using a relaxation labelling process. In this process, the matching line equation between the two images is used as a constraint function to remove the ambiguous labels from the two relational tables. This process is applied iteratively until two isomorphic relational tables are deduced. Since the labels in the two isomorphic tables are in one-to-one correspondence, the problem of matching the two images is reduced to a problem of matching regions to regions, edge segments to edge segments, and vertices to vertices. Again using the matching line equation as a geometric constraint function, the corresponding points in the two images may be searched in the corresponding edge segments. The three-dimensional object geometry is then computed using the matched corresponding points.