Automatic video segmentation using a novel background model

Fully exploits both temporal information from background and spatial information from foreground. In the proposed technique, the background regions of one scene are first warped into a large sprite image, which includes all visible parts of background throughout the sequence. This is completed automatically from raw sequence with the assistance of rough motion segmentation. The initial foreground regions are detected in each frame by subtracting the background derived from the sprite image, and then divided into many homogenous texture regions by watershed algorithm. Further merging and refining eventually determines the boundaries of foreground. The major contribution of this paper is to organize temporal information of sequence by virtue of sprite technique. Since background sprite image is formed with multi-parameter motion model across the sequence, the proposed scheme is robust rather than the method simply utilizing differential information among adjacent frames. Furthermore, the proposed scheme allows video sequence with various global motions, such as shifting, rotation and zoom.