Author:
Niranjan Mulay
Affiliation:
Sasken Communication Technologies Ltd, Products Division, India
Keyword(s):
Video compression, fast motion estimation, block matching, direction biased search patterns.
Related
Ontology
Subjects/Areas/Topics:
Image and Video Processing, Compression and Segmentation
;
Multimedia
;
Multimedia Signal Processing
;
Telecommunications
Abstract:
Motion estimation (ME) is computationally the most challenging part of the video encoding process. It has a direct impact on speed and qualitative performance of the encoder. Consequently, many sub-optimal but faster ME algorithms have been developed till date. In particular, the Three Step Search (TSS) and Four Step Search (FSS) algorithms have become popular because of their ease of implementation. The TSS algorithm is a uniformly spaced block matching algorithm, which performs better in case of large motion. On the other hand, the New Three Step Search (NTSS) and FSS are center-biased algorithms that outperform TSS in case of smooth correlated motion. Later, another center-biased search technique namely, the Diamond Search (DS) algorithm was introduced which was proved to deliver a faster convergence than FSS in case of smooth motion scenarios. However, the performance of the center-biased algorithms degrades in sequences having consistently large or uncorrelated motion as they be
come susceptible to getting trapped in local minima near the center. In this paper, two novel ME algorithms, namely, dual square search (DSS) and dual diamond search (DDS) are proposed in order to strike a balance between the center-biased and uniformly spaced search techniques. The proposed algorithms suggest that a decision to shift the search center should be delayed till the candidates on a coarse as well as fine grid are evaluated. Moreover, these algorithms are modeled to exploit motion vector distribution found in most of the real world video sequences by giving more precedence to candidates near the center, followed by the candidates in the horizontal and vertical directions than those in the diagonal direction. The performance of the proposed algorithms is compared with TSS and FSS algorithms in terms of computational speed, motion compensation error and the compression achieved for various kinds of video sequences. The tested sequences show that both these algorithms can be substantially faster than TSS and FSS. The proposed ME algorithms promise to achieve a balanced tradeoff amongst ‘speed - bit rate - quality’ for different kinds of motion sequences.
(More)