Memory access reduction in motion estimation

Abstract

In a video encoding system, motion estimation plays a key role to improve coding efficiency. In the motion estimation, the large amount of external memory access results in large power consumption and performance degradation. The conventional search area reuse algorithm reuses data of search area in on-chip memory by a fixed search center so that it can reduce the amount of external memory access. However, it causes the large coding efficiency degradation in fast motion video sequence. Moreover, in multiple reference frame motion estimation (MRF-ME), this algorithm requires large size of on-chip memory to store data of search areas in all reference frames. In this thesis, we propose the selective search area reuse (SSAR) algorithm which reuses search area only if the change of motion vectors is small. The algorithm not only reduces the amount of memory access like the conventional reuse algorithm, but also prevents coding efficiency degradation by tracking real motions. In addition, we propose the frame-level dynamic search range algorithm based on the SSAR algorithm. The number of MBs, which have a larger motion vector than a search range, is utilized as the condition to increase or decrease a search range. Then, we extend our algorithm to MRF-ME by changing the search order of reference frames and the search center in it. This algorithm can gradually reduce the amount of memory access according to the number of on-chip memories. Finally, we propose the SSAR algorithm in MRF-ME with on-chip memory usage switching method based on the dynamic search range algorithm. In this algorithm, we make use of the unused region of the on-chip memory as multiple on-chip memories for search area in multiple reference frames when a search range is small. By using these algorithms, we prevented quality degradation of 0.78dB with 10.14% memory access overhead in fast motion video sequence. And, we achieved memory access reduction between 27.79% and 54.71% depending on the numb...