A Video Mosaic Algorithm for Compressed Video

A video mosaic method is proposed in this paper. Firstly, phase correlation is used to roughly compute translation offset between the first corresponding frames of two video streams, which speeds up corner match procedure and improves matching stability as well. Secondly, SIFT method is used to detect and match corners in the overlapped region. Then, a RANSAC technique is applied 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. For the other frames left, we calculate a global motion vector between consecutive frames from motion vectors included in the compressed video data, thereafter obtain projection matrix between two frames from that of the previous frame together with the global motion of each input video sequence. In the last step 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 experiments show that the proposed algorithm exceeds existing ones at matching speed and stability.