Feature Matching of Unrestricted Scenes Based on Topology and Triangulation

Repeated contents and non-single-plane scenes usually lead to serious mismatching or missed matching. A novel feature matching method using topology structure and triangulation is proposed to match images without any constraint conditions. First, using Euclidean distance to measure similarity, similar features are significantly closer than dissimilar contents. K nearest neighbors distance ratio algorithm is proposed to find matching candidates which have significantly smaller distances. Second, 1-to-1-matches in reference image are triangulated, and the triangulation is mapping to the target image. The outliers which do not conform to the triangulation rules are moved. Then, all feature-sets are hierarchically triangulated based on topology similarity measure, and m-to-n-matches are reduced to 1-to-1-matches. The experimental results show that with the proposed method, mismatching and missed matching can be greatly reduced.