基于三维捕获数据的人体运动分析关键技术研究

发布时间：2018-03-03 11:13

本文选题：运动捕获数据　切入点：流形降维　出处：《浙江大学》2007年博士论文　论文类型：学位论文

【摘要】： 二十世纪九十年代以来，随着运动捕获技术的兴起，以及设备技术的进步，大量的三维人体运动捕获数据生成，并被广泛的应用在计算机动画，游戏，医学仿真，电影特技等领域。在此背景下，基于运动捕获数据的研究，已经成为近年来计算机视觉，图形学研究领域和动漫应用领域的一大热点。随着大量三维人体运动捕获数据库的出现，使得如何对人体运动数据进行高效分析和处理，从而有效利用运动捕获数据库成为一个新的挑战。在这些挑战中，如何从复杂的人体运动中找到能正确表达运动信息的本质特征，如何从浩如烟海的运动数据中检索出用户所需要的运动，以及如何去合理高效的重用处理后的运动数据，都是该领域关心的热点问题。本论文的目的是针对基于三维捕获数据的人体运动分析处理技术的上述关键问题，研究如何正确的提取人体运动的特征，如何不失真的对运动数据进行降维以及索引，探索如何用机器学习的方法对运动库的数据进行自动分析，以及如何结合子空间的方法对人体运动进行编辑和合成，从而实现利用运动捕获数据库进行动漫创作，电影特技制作等实际应用。主要完成了以下工作： 1．运动捕获数据的特征提取。不同于传统的运动特征提取方法，，考虑运动的几何结构能够较真实地反映运动内在属性。二维的几何特征只能较好的表达运动的局部几何结构，本文提出的三维时空特征则从全局几何结构角度出发，不仅能够有效表达各个关节点独立运动信息，而且能够反映关节点之间相互作用的运动属性。 2．运动的特征数据的降维和分段。提取出来运动特征维数较高，为了克服维数灾难，实现运动数据的识别和检索，本文利用主成分分析(Principal Components Analysis，PCA)和扩展等距特征映射(Isometric Feature Mapping，ISOMAP)方法对高维运动特征进行线性和非线性降维。同时，对于长运动序列，还利用上述方法对运动数据在子空间进行分段，将其分解为单个运动类型的短运动片段，以利于后续分析和处理。 3．运动数据的高效索引和快速匹配。为了适应大规模数据库的需要，本文对降维分段等预处理之后的运动数据提出了基于决策树，多示例学习和集成学习的有效索引方法以过滤绝大部分和查询例子不相关的运动样例，综合考虑人体各个关节点对运动影响的重要性，达到加速检索的目的。 4．基于子空间的运动编辑和合成。由于运动数据在子空间的内在隐藏属性能够在子空间被反映出来，本文提出了对运动数据在流形子空间进行线性和非线性映射，从而分离出风格参数的方法，进而对其进行编辑生成新风格运动。 5．系统实现。本文实现了上述运动捕获数据的分析处理技术和方法，并给出了人体运动检索和运动编辑合成的结果。
[Abstract]:Since 1990s, with the development of motion capture technology, equipment and technology, a large number of 3D human motion capture data generation, and is widely used in computer animation, games, medical simulation, movie stunt and other fields. In this context, the study of motion capture data based on computer vision in recent years has become a major. Hot graphics and animation applications.
With the emergence of a large number of 3D human motion capture database, so how to conduct efficient analysis and processing of human motion data, thereby effectively using motion capture data has become a challenge. In these challenges, how to find the essence can correctly express the motion information from complex human motion, how to retrieve the user needs the movement from motion data in multitude, and how to efficiently reuse treatment after the motion data is a hot issue in the field of concern.
The purpose of this thesis is aiming at the key problems of 3D human motion capture data processing technology based on feature extraction of motion to correct research, how to reduce the dimensionality of the distortion and the index of motion data, explore how to use the machine learning method of motion database for automatic data analysis, and the method of how to combine subspace editing and synthesis of human motion, thus using motion capture data in animation creation, application of film stunt. Mainly completed the following work:
The 1. motion capture data feature extraction. Motion feature extraction method is different from the traditional structure, considering the geometric motion which can reflect the intrinsic property of motion. Expression of local geometric structure movement features of two-dimensional better, three-dimensional space-time is proposed in this paper features from the perspective of global geometric structure, not only can effectively express each joint independent motion information, and can reflect the interaction between joint movement attributes.
Dimensionality reduction and feature data of 2. motion segment of the extracted motion. The higher dimension, in order to overcome the curse of dimensionality, recognition and retrieval of motion data, using principal component analysis (Principal Components, Analysis, PCA) and extended Isometric Feature Mapping (Isometric Feature Mapping, ISOMAP) method for high dimensional linear and nonlinear motion features dimensionality reduction. At the same time, for long sequences using the above methods, also to segment motion data in the subspace, which can be decomposed into short fragments of a single motion motion type, in order to facilitate the subsequent analysis and processing.
3. efficient indexing motion data and fast matching. In order to meet the needs of large-scale database, this paper on the motion data preprocessing after dimensionality reduction segmentation is proposed based on the decision tree, the effective index method of multi instance learning and ensemble learning to filter most of sample movement and query examples related to the importance of considering the influence of human body the joint exercise, to speed up retrieval.
4. subspace based motion synthesis and editing. Due to the movement of data in the subspace of the intrinsic properties can be reflected in the sub space, this paper puts forward to motion data of linear and nonlinear mapping in manifold subspace, thus separating methods of style parameters, and then edit it and generate a new style of sports.
5. system realization. In this paper, the analysis and processing techniques and methods of the above motion capture data are realized, and a person is given.
The results of body motion retrieval and motion editing synthesis.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2007
【分类号】：TP391.41

【引证文献】