摘要
针对已有的运动捕获数据关键帧提取方法常常忽略运动数据局部拓扑结构特性问题,提出了一种基于拉普拉斯分值LS特征选择的人体运动数据关键帧提取方法。该方法首先从原始运动数据集中提取两种代表性的特征向量并对其归一化,利用LS算法对组合后的特征向量进行打分和特征权重学习,以获取能够判别性揭示局部运动信息的特征子向量;其次,通过构建综合特征函数并基于极值判别原理,得到初始候选关键帧序列;最后,根据时间阈值约束和姿态相似判别策略,利用改进的k-means算法对候选帧进行聚类筛选,以达到去除冗余关键帧的目的,从而得到最终关键帧序列集合。仿真实验结果表明,该方法提取的关键帧序列具有典型性,能较好地对整体运动捕获数据进行视觉概括。
Existing key frame extraction methods often fail to reveal the local topological structure of motion capture data. To this effect,we present a Laplacian Score （LS） based feature selection approach to extract the key frames from the motion capture data. The proposed approach first extracts two kinds of representative and normalized feature vectors from the original motion capture data,and then employs LS algorithm to learn the scores and weights of the combined feature vectors. Accordingly,the discriminative feature subspace holding the promise of identifying the local motion information can be constructed. Subsequently, the initial key frame sequences can be obtained by the utilization of the comprehensive characteristic function and the extreme discrimination principle. With the constraints of the time threshold and discrimination strategies of similar poses,we further utilize the improved k-means algorithm to cluster the candidate frames such that the final key frames can be obtained to remove the redundant ones. The experimental results show that the typical key frames extracted by the proposed approach have better visual summary of the whole motion capture data.
出处
《计算机工程与科学》
CSCD
北大核心
2015年第2期365-371,共7页
Computer Engineering & Science
基金
国家自然科学基金资助项目(61202298
61202297
61300138)
关键词
关键帧提取
运动捕获
局部拓扑结构
拉普拉斯分值
特征选择
key frame extraction
motion capture data
local topological structure
Laplacian Score
feature selection
