| Video images are motion images compared with still images.It is composed of multiple images and every image is called a frame.When these frames (still images) continuous play and the average speed is more than 24 per second, we can find theses images become a motion video.A video mosaic method is proposed in this paper. The experiments show that the proposed algorithm exceeds existing ones at matching speed, stability. Video is composed of image sequences. Video Registration is a technique that relates or aligns different images taken from different viewpoints. Video mosaic technology uses the result of video registration to build large view panorama-like image from small adjacent image series, which could enlarge the user's view scope and increase image resolution as well.Till now, video registration and mosaic technology is widely used in the field of virtual reality(VR),video compression, image super-resolution, intelligent surveillance system, etc, which came to be an active research area in computer vision during recent years.We can regard video registration as multiple images registration.But we can not do it completely. Video imagery has many advantages over still frame imagery. For example, it provides context and timing relationships, which are suitable for dynamic scene monitoring and action verification. However, manipulation of video requires automatic processing and analysis, vast amounts of storage, efficient search methods, high bandwidth communication, and real-time implementations. Video sequences also have a lot of redundancy because of the large overlap between consecutive frames.Feature-based, direct pixel difference optimization based, and Fourier based method are three typical ways of image registration, which have its own appropriate application area separately. The key problems focus on how to increase the algorithm speed, to increase the registration precision and to enhance the robustness of these registration methods.This paper did research on video registration and mosaicing technique and presented some new effective algorithms based on point-matching.Consider that most of the frames in compressed video streams of typical video coding standards such as MPEG-2 and H.264/AVC are usually motion-compensated. Their motion information can be applied to global motion tracking in the video streams. So we propose a fast video stitching method using compressed-domain information in video streams.Firstly, phase correlation is used to roughly compute the translation offset between the first corresponding frames of two video streams, which speeds up corner match procedure and improves matching stability as well.Secondly, in the overlapped region SIFT method is used to detect corners and register them. Then, RANSAC algorithm is used to eliminate outliers to ensure effectiveness of the matched corner pairs. Singular Value Decomposition-Least Square(SVDLS) method and Levenberg- Marquardt optimization are used to robustly determine the 8 parameters transform model.For the other frames, we calculate global motion vector between consecutive frames from motion vectors included in compressed video data and obtain projection matrix between two frames from the projection matrix of the previous frame and global motion of each input video sequence.At the last of the algorithm, a multi-band blending technique is used to generate the final panorama. Invalid parameters are verified by the translation offset to make Levenberg-Marquardt optimization more successful.The algorithms in this paper are testified to be effective by lots of interesting, promising experimental results too. |
PDF Full Text Request