(A) fast multi-resolution block matching algorithm for multiple frame motion estimation

Abstract

Video compression plays an important role in transmission and storage of digital video data. The main idea to achieve compression is to remove temporal and spatial redundancies existing in video sequences. One effective method commonly used in reducing temporal redundancy is motion-compensated predictive coding. For motion-compensated coding, block matching algorithms (BMA) have been widely used as a motion estimation technique in most video coding systems due to its simplicity and effective compromise between prediction quality and motion information overhead. To improve the coding performance, multiple frame motion estimation (MFME) has been recently developed to extend the motion search range by using multiple decoded frames as the reference, instead of using only one decoded frame as in the conventional motion compensation scheme. It provides a high coding performance, however, its computational complexity is significantly increased as compared to the single frame motion estimation (SFME). In this thesis, we propose a fast block matching algorithm for MFME. Its major objective is to perform a fast search while maintaining a PSNR performance similar to a full search block matching algorithm (FSBMA). For reduction of computational complexity, we adopt a multi-resolution search with three resolution levels. Then, we adaptively modify the framework of the multi-resolution search according to the type of motion vector (MV), which is determined by using the MV information of neighboring macroblocks (NMBs). We also effectively reduce the number of MV candidates at the original resolution level by utilizing the sum of absolute difference (SAD) previously obtained. In computer simulation, it is proved that the proposed algorithm provides a similar PSNR performance with dramatically reduced computational complexity as compared to the FSBMA. In particular, for sequences with fast motion or frame skipping, the proposed method provides more prominent PSNR performances w...