Video Scene Change Detection Based On Image Feature Analysis

With the rapid development of computer technology,70% of the information comes from the image, which can bring people intuitive and simple experience. With the development of image coding and display technology, types of information to the people’s demand become diversified. And people hope to be able to communicate real-timely through various ways of voice and video. Therefore, the requirements of improving the quality of video information are increasing constantly.At present the video is divided into two categories:two-dimensional video and three-dimensional video. In recent years, with advances in 3D movies at home and abroad, people are no longer satisfied with 2D video. With its unique experience and the realistic effect of viewing, stereoscopic video attracts more and more people into the cinema to feel the visual impact brought by it. Thus it gives a strong impetus to the development of the stereo video and multi-view video.The application of 3D video involves many fields, such as three-dimensional television (3DTV), video phone, video surveillance, medical and other fields.3D video data is extremely large compared with the 2D video, so the stereo video data transmitted through limited bandwidth of the network become a difficult problem.3D video needs to simultaneously transmit multiple viewpoints of video data and corresponding depth information. However the current network bandwidth is hard to bear such a large 3D video coding data. Therefore, the compression of the original data, which can simultaneously adapts the load of the current network, and guarantees the visual perception of the 3D video, is a quite challenging task. Scene change detection technology is the basis of video analysis and retrieval. At the same time, the scene cut point detected by scene change detection technology can be considered as a basic unit in video coding. The amount of encoded video data can be reduced as much as possible without affecting the viewing effect by locking the key frames. Thus,2D/3D video scene change detection algorithms play an important role for the rapid development of image and video technology.This thesis includes mainly two parts:3D video scene mutation detection and 2D video scene change detection. Using the image features such as motion vectors, disparity vectors and SIFT feature points which can reflect the characteristic of 2D/3D video, the change of the 2D/3D video scene is detected by quantitative analysis.For 3D video scene mutation detection, this thesis jointly uses the motion vector of the single view and the disparity vector of the stereoscopic video to solve a stereoscopic video scene abrupt change detection problem. The proposed algorithm detects scene abrupt change using motion vectors and disparity vectors generated during multi-view video coding. Therefore, it can not only guarantee the timeliness of the algorithm, but also reduce computational complexity.For 2D video scenes mutation detection, this thesis proposes a scene mutation change detection algorithm combined with SIFT (Scale Invariant Feature Transformation) feature point extraction. The proposed algorithm can judge scene change during image matching. Therefore, the algorithm can not only be applied widely, but also guarantee the accuracy of scene change detection. Meanwhile, the dense SIFT feature point parameters combined with the hierarchical clustering algorithm are used for video segmentation pretreatment. After that,2D video scene gradual change detection is done by using gray histogram in each cluster, which improves the accuracy of 2D video scene gradual change detection.