Image Feature Extraction And Matching Algorithm

This thesis is supported by the "intelligent video surveillance system" project, which is applied to monitor the region for the unusual circumstances of people and things and timely warning, so as to achieve real-time monitoring purposes. By the caparison of the virtues and the shortcomings of the key technology in the various processes, the appropriate and effective methods are chosen to availably improve the intelligent video surveillance in terms of the system requirements so that it can help monitoring person analysis the monitor screen, filter redundant information, obtain useful information from the image in time, and finally accomplish the intelligent control.This thesis aimed at the basic theories and key technologies of the image feature extracting and matching.The local feature point's vector was chosen as the image property descriptor for image-matching. By analyzing five classical operators of SUSAN, Moravec, Harris, Forstner and SIFT (Scale Invariant Feature Transform), the SIFT algorithm had a better robustness in brightness changing, zoom scaling and rotation, noise and affine transformation. The feature descriptor of SIFT had abundant information, good independence, stable algorithm and high recognition rate, it can meet the system requirements. Thus the SIFT algorithm was studied and improved in this thesis.Recognition rate and the stability of the SIFT algorithm is superior to other algorithms, but it has the shortcoming of high calculation, high complexity and poor real-time performance. SIFT algorithm mainly includes four steps such as establishment of scale space, detection of extreme space points, determination of feature points direction and generation of feature descriptor. By getting the running time of each step, it was found that the time for the generation of descriptor features accounts for more than 70 percent of the whole time. The SIFT algorithm was improved in this thesis in the following aspects. In the feature extracting part, the number of dimensions of a feature descriptor is reduced from 128 to 48, omitting the determination of primary direction. The dimension performance also has been analyzed and the reasonableness of selected dimension has been verified. In the feature-matching part, in order to further remove the false matching points and improve matching rate, second-matching is added after initial. It is proved from the tests that the improved SIFT algorithm can significantly increase"real-time"and matching accuracy.