A Study Of Image Sequence Motion Estimation Techniques Based On Fast Block-Matching

As the key technique of computer vision and digital video processing, motion estimation of sequence images shows more and more important position and role not to be ignored in the various directions of production and life in society and has wide application value in industry, commerce, country defense, medicine, internet technology, etc. Recently, due to its low computational complexity and high compression efficiency, block matching technique has been adopted by many video coding standards such as ISO MPEG-1/2 and ITU-T H.261/263, and gets great improvements and rapid developments in methodology and application and becomes a new research hot point of digital video technique. On the base of comprehensive summarization and deep analysis for current block matching technique research work, a block matching algorithm based on motion vector distribution has been proposed in this thesis and has been applied to hierarchical block-matching algorithm based on the wavelet transform. Experimental results show the new algorithm has obvious advantage in respect of enhencing accuracy and reducing the computational complexity, etc.At the beginning, a fully survey about motion estimation algorithms is given. Matching algorithm based on block becomes to be considered, owing to its simplistic method, good estimation efficiency and potential to be implemented for the real-time application using the available hardware. On the base of understanding for block matching problem in nature, a systematic summarize is made for all current block matching algorithm in the world. From the different search algorithm and search criteria adopted by two-dimension logarithmic search algorithm, cross search algorithm, gradient descend search algorithm based on block, and diamond search algorithm, etc, we have analysed their applicability, and compared them in terms of computational complexity.In the aspect of motion estimation based on motion vector distribution characteristics, we firstly introduce the motion vector distribution characteristics of sequence images, i.e., center biased distribution characteristics of motion vector, the correlation among the motions of adjacent blocks and the direction of SAD value. This is basic of this thesis. Then the analysis is made to the current start search point prediction method, matching criterion. Taking good advantaging of these ideas, a developing algorithm, start search point prediction method based on motion correspondence and motion estimation matching criterion based on vision feature, is proposed. Aiming to the computational redundancy which is made by the current search mask and search strategy using in micro-block motion estimation, a new DSM mask which combines the diamond and square mask and a new search strategy which is along the descend of the SAD value is proposed . Taking good advantaging of these criteria, method, search mask and search strategy, a new block matching algorithm, fast block matching LCB algorithm, which is based on motion vector distribution characteristics, is exploited. At last, the computational complexity and accuracy of the solution is compared with existing algorithms. Experiment result shows the good quality performance, efficient and robust of the new algorithm.A improved method of hierarchical block matching motion estimation based on the wavelet transform is presented. Firstly, the relation of the wavelet transform and the multiscale is mentioned and a concision introduction of 2D wavelet decompose and reconstruction algorithm and multiscale discription of 2D image are made. Hierarchicalmotion estimation theory is the application of multiscale theory and the wavelet theory in the sequence image analysis. It used multiscale feature of the wavelet to get sub-images which own different resolution. Then block matching in the different scale is done. The developed method firstly uses the motion estimation algorithm based on motion vector distribution characteristic which is mentioned in the chapter 3 to do block matching in the low resolution, then does more precise estimat...