CONCLUSION

    The search window of the DS algorithm is limited due to the size and boundary is fixed . And it may leads to distortion if the checking points are outside the search window . However , it can be improved by larger the search window if necessary . The size of the search window is not restricted , but the computational cost should be taken into consideration .
    The reasons that I don’t use full search ( FS ) algorithm to estimate the motion is the computational time and the motion of the video . Because there’s no unstable shaking in the testing video , the changes of the positions between adjacent frames in a video sequence aren’t many . If let the FS algorithm exhaustively tests all the blocks within a frame , it not only wastes time , but also may mislead the search path to a wrong direction ( a large area in the same color and influenced by the shadow , Figure C-1)

    The motion detection of foreground and background in the video is already established . After that , I’m going to extract the foreground object . The system estates the motion by separate frames into 16*16 motion blocks . It’s possible for each block to cover both background and foreground , therefore if the system decides which block belongs to foreground by its motion , and takes a motion block as unit to extract foreground , the output can’t be perfect .
    Next step of my implementation is combined with watershed algorithm . Another group of this lab applies watershed algorithm to deal with the extraction of foreground from picture . We’re going to cooperate to extract foreground from videos . Similar to the methods mentioned in the paper “ Lazy Snapping “ , the system will ask users to do a quick object marking steps in first frame and last frame to indicate foreground and background by lines (Figure C-2). We take the small regions after the watershed segmentation as unit to extract foreground . Other frames in the video will depend on the data of colors and motions of foreground and background to extract foreground object automatically .