Feature Points Matching Based on Geometric Constraints

Feature matching is a basic problem in computer vision, and the method based on feature point is the most common one which has important research value. There are many factors which may affect the matching result of the feature points. In this paper, a novel matching method combined texture information and geometric constraints under projective transformation is proposed. First, the newly developed projective invariant "characteristic number" is introduced to compute the geometric descriptors, then histograms of characteristic numbers for each point are built and the Bhattacharyya distance is used to measure the geometric similarities. Finally, the geometric constraints are applied to the descriptors based on local texture information to generate the criteria of points matching. Experiment results show this method can improve the matching accuracy effectively. It also performs well against large viewpoint changes and senses with similar textures.