结合HMM隐状态基元和贝叶斯准则的运动捕获片段过渡

Capture Data Transition by Combining HMM-based Hidden Primitives and Bayes Rule

针对运动捕获序列拼接时常常忽略其运动语义自然过渡问题,提出了一种结合隐马尔可夫模型隐状态基元和贝叶斯准则的运动捕获数据片段过渡方法.首先,提取两种代表性的人体骨架特征并归一化,得到组合特征数据矩阵来表示原始运动数据;其次,采用HMM方法对组合特征矩阵进行运动隐状态基元预测,发现其运动隐状态变化规律;紧接着,依据连接处不同运动序列隐状态预测结果,结合贝叶斯规则搜索运动片段作为连接片段;最后,对序列连接处进行四元数插值平滑处理,达到运动片段自然过渡的目的.实验结果表明,本文提出的方法能够较好地保持运动状态的自然过渡,且过渡区域无明显地拼接痕迹,有助于不同语义运动片段自动拼接时运动姿态的自然衔接.

Existing motion assemble methods often ignore the naturality of motion transitions. To this effect,this paper presents an efficient motion transition approach by combining the HM M-based hidden primitives and Bayes rule. First,the proposed approach extracts tw o representative feature vectors and normalizes them into an assembled feature matrix,featuring on briefly representing the raw data.Then,the typical HM M is employed to estimate the hidden primitives and simultaneously find their transitions. Subsequently,according to the pre-estimated result,the Bayes rule is adopted to search the proper motion clips for motion linking. Finally,quaternion interpolation is further utilized to smooth the connected motion clip,w hereby the w hole motion sequence can be w ell transited. The experimental results have show n that the proposed approach is able to keep the transition motion w ithout splicing trace and the assembled heterogeneous semantic motion clips are connected naturally.