Visual Motion Analysis And Feature Extraction

Content Based Video Retrieval (CBVR) is a newly arisen research directionafter Content Based Image Retrieval (CBIR), which provides new ways toretrieve video data that has similar visual content. As a unique type of informationthat discriminates from other media, motion has its special importance to theperformance of a video retrieval system. With an emphasis on the extraction ofmotion information and the representation of motion features, this thesis proposesa suit of solution to extract motion features for content based video retrieval. Thework of this thesis is a part of the Project Object Oriented Multimedia DatabaseApplications in Government-affair Information Processing, which grant from theScience Development Plans of Jilin province. The work of this thesis is mainly summarized as follows: 1. In this thesis, Motion information is divided into two categories: globalmotion and local motion. For global motion, this thesis gets block based motionvector field by Block Matching Motion Estimation (BMME), then an estimationmethod is taken to extract the motion model parameters. For local motion, thisthesis uses the Horn-Schunck method to extract the dense motion vector field. Themain steps are shown as follows: (1) Use a BMME method between the current frame and the reference frameto get the motion vector estimators of the frame pair. (2) An erroneous data rejection process is performed to get the reliable globalmotion vector estimators. (3) A global motion model is established (the bilinear model is chosen in thisthesis), and the model parameters are estimated from all the unrejected blocks'motion vector at step 2. (4) Reconstruct the global motion vector at the center of the block with themodel parameters estimated at step 3. (5) Calculate the difference between the block match's result and thereconstructed motion vector for all the blocks. If the difference of a block is largeenough, then mark this block as an inner of a local motion area. - IV -吉林大学硕士学位论文 (6) In every local motion area, use the Horn-Schunck method to extract thedense vector field. 2. In the MMBE process, based on the analyses of video data characteristicsand study of several other well-known fast algorithms, this thesis proposed anadaptive fast search algorithm, which benefits from the temporal and spatialcorrelations exist between frames of video sequences. In this algorithm, manyprimary improvements have been introduced, including: (1) Reliable prediction of initial search point There exist high correlations between motion vectors those in temporally andspatially adjacent blocks, and that can be seen from the vectors' value anddirection, which implies that a block's motion vector can be predicted with itsadjacent blocks' motion vectors. We can choose a predictor that reflects thecurrent block motion trend as the initial search position, instead of the (0,0)position, the predictor is more close to the global minimum position, and thatimproves the performance of the center based block matching algorithm both interms of speed up and veracity. In this thesis, we define the (0,0) motion vector(MV0), the motion vectors of the blocks on the left (MV1), top (MV2), andtop-right (MV3), and the motion vector of the collocated block in the previousframe at time t-1 (MVt-1) as Candidate Set A. In the first step of the search process,the best candidate is selected as the predictor from set A, and then the predictor istaken as the center of the search to carry on the following search steps. (2) Adaptive search modes for different motion type Among varieties of search patterns, the Diamond Search (DS) one has itsspecial advantages indeed, so we introduced the two patterns: Large DiamondSearch Pattern (LDSP) and Small Diamond Search Pattern (SDSP), and LDSPwas used to fast position in a wide scope, while SDSP was used to fine adjustwithin a...