A Computationally Adaptive Block-Matching Motion Estimation Algorithm VSK Reddy 1 and Somnath Sengupta 2 1,2 Department of Electronics and Electrical Communication Engineering Indian Institute of Technology Kharagpur, West Bengal, India-721302 Email: vskreddy2003@yahoo.com Abstract This paper describes a computationally adaptive motion estimation algorithm by exploiting the spatial-temporal motion correlation characteristics of the video sequence. In our approach, a set of initial candidate MV’s include center block and four neighboring blocks at the same location from the previous frame are used to track the motion of the current block. Among the set of initial MV candidates, the one with the smallest MAD is chosen as the reference MV candidate V R of the current block. We use the spatial correlation property between zero motion vector and reference vector V R for finding the best candidate MV and viewed as a starting point (i.e., new search center) for MV searching process. The proposed approach has been successfully tested with integer, as well as half-pixel accuracy and the results are encouraging. The experimental results reveal that the proposed algorithm saves up to 95% of computations, as compared to FSBM and the estimation accuracy is very close to that of FSBM. 1. Introduction Video compression plays an important role for digital video applications in recent years. Motion estimation is used to remove temporal redundancy by searching the best-matched block in the previously encoded frame. The simplest and most effective method of motion estimation is to exhaustively compare each N x N macro-block of the current frame with all the candidates blocks in the search range defined with in the previous processed frame and find the best matching position with the lowest distortion, this is called full search block matching algorithm (FSBM). The distortion measure is usually the Mean of Absolute Difference (MAD) for its simplicity, in which the candidate block with the minimum amount of distortion is considered as the best match. The distortion (D) for the candidate block at the displaced position (u 1 ,u 2 ) and the motion vector v are given by ( ) ( ) ( ) ∑∑ − = − = + + − + + + + = 1 0 2 1 1 0 2 1 2 1 , , , , , N k N l j i j i l x k x r l u x k u x s x x u u D (1) ) , , , ( min arg 2 1 2 , 1 2 1 x x u u D v p u u p ≤ ≤ − = (2) Where, N X N is the block size, (x 1 , x 2 ) is the top-left pixel position of the reference block with respect to the frame coordinates; r(i, j) and s(i+u 1 , j+u 2 ) are the luminance values of the reference block and the candidate block respectively. A full search block-matching process with a search range p has a search window of size (2p+N) x (2p+N) pixels and a total of (2p+1) x (2p+1) candidate blocks in the reference frame for each block of the current frame. The distortion values are computed for each of the candidate blocks and its minimum value is found from the set of (2p+1) 2 candidate blocks. Although the full search algorithm can indeed obtain the global optimal result, however this method has enormous computational complexity. In order to reduce the computational complexity of the motion estimation, several fast block- matching algorithms, such as three step search [1], the novel four-step search algorithm [2], the hexagon-based search algorithm [3], block-based gradient descent search algorithm [4] etc., have been proposed. All these algorithms are not optimal in the sense that instead of exhaustive search, only some fixed positions are searched, based on the predictions of motion. To achieve a best trade-off between the computational complexity of FSBM and degraded PSNR of motion compensated frame using faster algorithms, recently some researchers have investigated reduction of computational complexities of FSBM, without causing any significant reduction of PSNR. Haung et al [5] presented a predictive line search, based on the mean value of motion vectors of the causal neighbor macro-blocks. Chung et al [6] presented a new predictive search area approach, in which predicted search area can be obtained from sub areas of the neighboring blocks. Another direction for MV estimation approach is to exploit information from adjacent blocks by using spatial and temporal correlations of MV’s [7]. The main idea is to select a set of initial MV candidates from spatially or temporally neighboring blocks and chooses the best one as the initial estimate for further MV refinement. The initial estimate can be obtained using autoregressive model and only one candidate is chosen for estimation. The MV refinement process involves a full search still requires a considerable amount of computation. Even though many fast MV estimation techniques have been proposed as reviewed before, we fell that the spatial and temporal correlations of MV’s have not yet been fully exploited in reducing the search time while maintaining reasonable distortion trade-off. With these observations,