Image Matching And Application Based On Local Feature And Delaunay Meshes

Image feature matching technology is a major research content in computer vision, image processing and multimedia information retrieval. The essence of it is to look for one or more distance measurement methods in high dimensional feature space, making two or more images from different times, different sensors or different perspectives of the same scene be consistent in space. Image matching based on feature has a small amount of calculation, fast speed and be robust to image distortion, noise, shade, however, the matching accuracy largely depends on the quality of the feature extraction, without the location of the feature points and characteristics in the dataset. To solve this problem, this paper presents the method of Delaunay triangulation subdivision of the convex point sets in the feature space to improve the matching accuracy. And apply the method to non-overlapping multi-camera pedestrian re-identification problem, promising results are obtained in the public VIPER dataset. The main work is as follows:(1) This paper introduces the background and significance of the image feature matching. We discuss the image feature matching technology development status and key issues. Spatial structure information is introduced into data sets. This paper present elaborate ideas, theories and methods of Delaunay triangulation division.(2)Image feature is always lack of space and position information, to solve this problem, a method combined of local features and spatial subdivision is proposed. By introducing space subdivision, building on the composition of the feature point Delaunay mesh, this method uses edge information as the similarity measure in the image matching, Including not only the local feature, but also gradation, image texture underlying global information. Meanwhile, the expression of spatial information is in topology instead of in geometry, which makes this method has stronger robustness to nonlinear distortion of the image.(3) The above method is applied to non-overlapping multi-camera pedestrian re-identification problem.In consideration of poor image quality, small target and posture changing in the images, We use MSER regional feature, which is more stable than point. A hierarchical matching strategy, which is consist of coarse and fine match, is used to improve the accuracy of matching, and achieves good experimental results.