Research On Video Retrieval Algorithms Based On Local Features

With the rapid development of Internet and multimedia technology and the widespread use of mobile equipment, springing up a lot of video,and it is important to find the video that people wanted quickly in massive video.Traditional video retrieval method which mainly based on text cannot meet the needs of social development.In order to solve these problems, in the 1990's,people began to research and develop Content-Based Video Retrieval(CBVR),which make best of color, texture, shape, and motion information to describe the Video.The key technology of content-based video retrieval system included Shot segmentation, key-frame extraction and feature extraction.In this paper, on the basis of the existing algorithm, the key technology are studied and improved, and the main work is as follows:(1)Summary the typical abrupt shot-boundary,Gradual shot-boundary detection and key-frame extraction. Describing and analysising the local feature extraction algorithm and SIFT algorithm is studied in detail. Feature matching search algorithms and the matching algorithm for purificationare introduced, and RANSAC algorithm is analyzed in detail.(2)Present a method of shot segmentation which combine the method of the motion detection and the method of the pixel difference in the two continuous pictures.And the new method could avoid shot segmentation mistakenly,in the condition of little slow shot segmentation speed.(3)Put forward a method of key-frame extraction based on the amount of exercise.When the video has small global motion,the traditional key-frame extraction which based on the boundary is used. When the video has medium or large global motion,different sampling interval is used, the video is sampled by different interval as key-frames.The method can solve the problem that the numbers of the key-frames is too small,and it help improve the success of video retrieval.(4)Present a local feature descriptor which is based on LBP, and the descriptor using a binary pattern to describe local area of a image, and the integral image is used to speed up the process of median filtering, thus the speed of generating descriptor is faster than SIFT descriptor, which is ten times faster than SIFT descriptor.Under the condition of fuzzy, variant size, different JPEG compression, the matching results are better than SIFT descriptor.But the feature descriptor is mainly used in video retrieval, in order to speed up the generation of feature description, therefore the scale invariance and rotation invariance was not taken into account.