基于余弦距离的人体运动数据行为分割算法  

Human behavior segmentation method based on cosine distance for motion data

作　　者：邢薇薇[1] 仝磊鸣 张毅[2] 任程[3] 

机构地区：[1]北京交通大学软件学院,北京100044 [2]北京交通大学计算机与信息技术学院,北京100044 [3]煤炭科学研究总院,北京100013

出　　处：《中南大学学报（自然科学版）》2014年第4期1128-1136,共9页Journal of Central South University:Science and Technology

基　　金：国家自然科学基金资助项目(61100143;61370128;61272353);教育部新世纪人才计划项目(NCET-13-0659);北京高校青年英才计划项目(YETP0583);高等学校基本科研业务费项目(2012JBM124)

摘　　要：对人体运动捕捉数据进行行为分割是人体运动数据分析与合成中的关键处理步骤,为此,提出一种新的人体运动数据行为分割算法。采用骨骼夹角直方图刻画人体运动统计特征,使用余弦相似度作为人体运动数据骨骼夹角直方图特征的相似性度量,实现对运动行为的自动分割。对于给定的人体运动捕捉序列,首先定义滑动比较窗口,计算当前窗口范围内运动序列前、后2部分骨骼夹角直方图统计特征的余弦相似度,然后通过在运动序列上滑动该窗口,获得运动序列的余弦相似度曲线,曲线最小值位置即为不同类型行为的分割点。在CMU人体运动捕捉数据库上进行数值实验。研究结果表明:本文算法能够实现对人体运动捕捉数据的自动行为分割;与广泛采用的基于PPCA的行为分割方法相比,本文算法具有良好的性能。Considering that the human motion segmentation is one important process for human motion data analysis and synthesis, a novel segmentation algorithm for motion capture data was proposed to segment motions into distinct behaviors. The method was based on the assumption that motions with the same type had similar histogram of angle between bones. Angle histogram was employed to represent the human motion, and cosine distance was utilized to measure the similarity of motions represented by angle histograms. A motion sequence was given, and a sliding window which moved from the starting frame to the end was firstly defined. In each window, the cosine similarity between the first half window and the second half was computed, and so a cosine distance curve with the motion was obtained. The minimums of the curve identified the cut between different types of motion behaviors. The present method was tested on the Carnegie Mellon Motion Capture database. The results show that the method can achieve the automatic segmentation for human motion capture data, and has good motion behavior segmentation performance compared with the classical PPCA based segmentation algorithm.

关 键 词：人体运动分割 运动捕捉数据 骨骼夹角 余弦距离 曲线简化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]