基于多特征的人体骨架运动检索

发布时间：2018-02-01 19:55

本文关键词： 基于内容多特征骨架运动检索　出处：《山东大学》2015年硕士论文　论文类型：学位论文

【摘要】：近几年来,随着三维游戏一系列的创作产品不断地兴起,计算机不仅在在文化创作(例如广告设计、电影创作、动画特效)、人机交互、游戏创作、广告娱乐等应用中发挥着不可替代的影响,还广泛应用于教育事业以及国防建设、卫星研发等科技领域。计算机图形学技术以及计算机软硬件的迅猛发展更是让计算机有能力开发出这些产品,并且提供了更加便捷的途径。随着大规模的三维人体运动数据库的不断建立,我们需要从复杂的人体运动序列中找到可以准确代表整个运动序列的属性描述符,需要对人体运动数据进行高效合理地分析与处理,以及检索出符合用户需求的目标运动序列,这些工作都是任重而道远的。我们针对基于多特征的人体骨架运动数据的检索提出了一种高效的解决方案。第一个主要的亮点在于我们利用不同的分步提取标准,从运动序列中提取并描述多种特征。另外,为了更加便利高效地进行特征匹配,我们通过主成分分析法和聚类分析法对特征描述符进行降维,并且利用多运动直方图来表示每种特征中的一个运动序列。最后,通过测度并排序查询序列和数据库中的目标序列的运动直方图的相似度,得到最终的检索结果。多次的对比实验表明我们提出的算法性能和效率较为突出。其中,我们工作的主要贡献在于以下四点：1、运动数据的多种特征提取。考虑到人体骨架的几何特征可以比较真实地反映出运动的本质特性,因此选取四种具有代表性的几何特征来更加准确地描述运动序列以提高检索的精确度。传统的二维几何特征提取只是显示三维人体骨架运动的局部几何特征,但是在我们提出的基于多特征的人体骨架运动数据的检索算法中提取基于三维空间的位置关系的特征是从全局的角度出发,来表示出人体运动的几何特征。这样既可以有效精准地显示出每个关节点自身的独立运动特征,又可以清楚明确地反映出各个关节点之间相互作用的运动特征。2、运动特征描述符的降维。由于我们提取出来的三维人体骨架运动特征描述符的维数很高,为了避免所谓的维数灾难问题,达到精准的运动序列查询和检索目的,本文通过主成分分析法和聚类分析法等技术对特征描述符进行降维来方便后续分析和处理,争取以最少的代价达到更高精度的特征匹配。3、运动数据的特征匹配。本文针对三维人体骨架运动数据进行降维、聚类分析等预处理之后,提出来利用运动直方图来对处理结果进行分析表示,也就是说可以通过计算每个类别出现的频率建立出运动直方图,并求得两两直方图之间的欧氏距离来对查询序列和数据库中目标序列进行匹配和检索。4、实验效果的评价测度。本文对查询序列和数据库中的每一个目标序列的运动直方图的相似度使用MAP和P@n的评价指标进行实验效果的度量,对检索结果的性能进行评判。
[Abstract]:In recent years, with the continuous rise of a series of creative products of 3D games, computers are not only in the cultural creation (such as advertising design, film creation, animation special effects, human-computer interaction, game creation. Advertising entertainment and other applications play an irreplaceable role, but also widely used in education and national defense construction. The rapid development of computer graphics technology and computer software and hardware makes the computer have the ability to develop these products. With the establishment of large-scale three-dimensional human motion database, we need to find attribute descriptors from complex human motion sequences that can accurately represent the entire motion sequence. It is necessary to analyze and process the human motion data efficiently and reasonably, and to retrieve the target motion sequence that meets the needs of users. We propose an efficient solution for the retrieval of human skeleton motion data based on multiple features. The first major highlight is that we use different step-by-step extraction criteria. . In addition, in order to improve the efficiency of feature matching, we reduce the dimension of feature descriptors by principal component analysis and clustering analysis. And the multi-motion histogram is used to represent a motion sequence in each feature. Finally, the similarity of the motion histogram between the query sequence and the target sequence in the database is measured and sorted. Many experiments show that the performance and efficiency of the proposed algorithm are outstanding. The main contribution of our work lies in the following four points: 1. Taking into account the geometric features of human skeleton, we can truly reflect the essential characteristics of motion. Therefore, four representative geometric features are selected to describe the motion sequence more accurately to improve the retrieval accuracy. The traditional two-dimensional geometric feature extraction is only the local geometric feature of the three-dimensional human skeleton motion. But the feature of position relation based on three-dimensional space is extracted from the global point of view in the retrieval algorithm of human skeleton motion data based on multi-feature proposed by us. In order to represent the geometric characteristics of human motion, it can effectively and accurately display the independent motion characteristics of each node itself. It can also clearly reflect the movement characteristics of the interaction between different nodes. 2, the dimension reduction of motion feature descriptor. Because we extracted the dimension of three-dimensional human skeleton motion descriptor is very high. In order to avoid the so-called dimension disaster, accurate motion sequence query and retrieval can be achieved. In this paper, the principal component analysis and clustering analysis are used to reduce the dimension of the feature descriptors to facilitate the subsequent analysis and processing, and to achieve a higher accuracy of feature matching at the least cost. Feature matching of motion data. After dimensionality reduction, clustering analysis and so on, the motion histogram is proposed to analyze and express the processing results. That is to say, the motion histogram can be established by calculating the frequency of each category, and the Euclidean distance between the two histograms can be obtained to match and retrieve the query sequence and the target sequence in the database. The similarity of the motion histogram of the query sequence and each target sequence in the database is measured by using the evaluation index of MAP and Pkinn. The performance of retrieval results was evaluated.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】