Abstract: Due to the quad-tree with nested multi-type tree coding tree architecture used in versatile video coding（VVC）, encoding process has great computational complexity, and this brings difficulties to practical use. To solve this problem, a fast coding unit（CU） partitioning algorithm based on temporal and spatial information is proposed. The proposed algorithm consists of two parts, in the first part, according to the shape of current CU, three different models are utilized to derive motion vectors from corresponding area in the reference frame, according to distribution of motion vectors in models, the CU partition modes that need to be traversed will be reduced by early selecting process; the second part utilizes the correlation of prediction modes between the current CU and adjacent same size CUs, by comparing their prediction modes and rate distortion cost, determines whether to do further traverse in the next depth, thus partition depths that need to be traversed is reduced. Experimental results show that, comparing with the latest encoder VTM10.0, the proposed method can save 39.28% of encoding time with 1.62%  increase in BDBR and 0.05 dB increase in PSNR in random access mode.

Key words: temporal and spatial information, coding unit, motion vector, prediction mode

摘要： 通用视频编码（VVC）使用的四叉树加多叉树的块划分结构使编码过程的运算复杂度极高，这给编码器的实际应用带来了困难。为了解决这一问题，提出了一种基于时空域信息的编码单元（coding unit，CU）快速划分算法。该算法分为两部分，第一部分根据当前编码CU的形状，使用三种不同的模型在参考帧对应区域提取运动矢量，根据模型中运动矢量的分布特性提前筛选划分模式，从而减少需要遍历的划分模式数量；第二部分利用相邻同尺寸CU的相关性，通过衡量当前CU和相邻CU的预测模式和率失真代价，决定是否进入下一深度进行遍历，从而减小所需遍历的深度。实验结果表明，在随机接入模式下，与官方测试平台VTM10.0相比，在平均输出比特率增加1.62%，平均峰值信噪比下降0.05?dB的代价下，编码时间减少了39.28%。

关键词: 时空域信息, 编码单元, 运动矢量, 预测模式