运动捕获数据较为广泛的应用于影视制作、游戏设计和体育专项训练等领域。随着运动捕获技术、运动合成技术等的飞速发展，高效地重用已有的运动捕获数据成为趋势。但是，在实际应用过程中，存在运动捕获数据序列过长等问题，这样限制了运动捕获数据在各领域的重用效率，同时也带来了存储不便等方面的问题。由于运动捕获数据分割和分类是重用数据，理解和构建运动模型的一个基础工作，所以我们希望实现运动捕获数据的合理分割并且进一步实现分类存储。近年来有许多科研团队围绕运动捕获数据的分割和分类开展了大量的工作，但是由于运动的多变性、不确定性和周期性等因素，准确且合理的分割和分类运动捕获数据仍然是一个巨大的挑战。 本文主要基于主题模型以及谱聚类在运动捕获数据分割和分类任务上的应用进行了相关可行性的分析。同时，在该任务上对主题模型和谱聚类等方法的融合算法进行了深入的研究。论文的主要研究工作内容如下： 1、根据运动捕获数据本身的特点，结合谱聚类算法的重要性质，本文分析了将谱聚类算法用于运动捕获数据分割和分类任务的优势和原因，提出了基于谱聚类的分割分类算法。首先进行数据预处理，并重新定义了适用于运动捕获数据中欧拉角度数据的距离度量方法。然后采用谱聚类对预处理后的运动数据进行初步分割，经过统计滤波之后得到初始分割结果。基于此粗糙结果，结合方差分析方法分段对原始运动数据进行自适应加权，目的是计算出在表征运动方面优于原始运动数据分段带权特征矩阵。再通过谱聚类对带权特征矩阵进行重分割聚类得到比初始分割更精确的结果。其中，基于方差分析的自适应加权过程也为其它分割分类算法提供了良好的数据初始化思路。 2、通过将运动捕获数据库中运动类型、姿态或者是运动片段的关系类比于语料库中主题和单词的关系，本文拟采用主题模型的思想来分析运动捕获数据。比较了运动数据库和语料库之间的异同，阐明了主题模型思路应用于运动捕获数据分割和分类任务的可能性和具体方法。并结合任务的实际情况选择引入稀疏算子的稀疏主题模型。在此基础之上，介绍了将稀疏主题模型和分类器相结合的有监督分割分类策略。同时，借助基于谱聚类的分割分类算法的思想，本文进一步提出无监督的基于稀疏主题模型的二次谱聚类（Double Spectral Clustering-Sparse Topic Model，DSC-STM）算法，DSC-STM是融合了谱聚类算法和稀疏主题模型的三层级联算法，通过谱聚类和方差分析对运动捕获数据进行加权初始化，再利用稀疏主题模型提取初始化后数据的稀疏运动-主题（m-topic）特征，最后，在稀疏m-topic特征上利用谱聚类进行重分割得出最终结果。m-topic特征既保持了原始运动数据的特点，又抽象出运动数据的类型信息，因此不仅具备稀疏的特性，更具有较强的表征运动类型信息的能力。 本文在CMU mocap database上进行运动捕获数据的分割和分类实验，并与其他算法进行比较。实验结果验证了本文所提分割分类算法的有效性和鲁棒性。

Motion capture data is widely used in film and television production, game design and sports training and so on. With the rapid development of motion capture technologies and motion synthesis technologies, it has become a trend to reuse existing motion capture data efficiently. However, there are some problems in practice, such as the too long sequences of motion capture data, which limits the reuse efficiency of motion capture data in various fields, and also brings problems such as in-convenience in storage. Since the segmentation and classification of motion capture data is a fundamental work for reusing data and understanding and building motion models, we hope to achieve reasonable segmentation of motion capture data and further classified storage. In recent years, many researchers have done a lot of work about the segmentation and classification of motion capture data, but due to variability, uncertainty and periodicity of motion, obtain the segmentation and classification of motion capture data accurately and reasonably is still a challenge. In this paper, the task of segmentation and classification of motion capture data based on Topic Model is studied in depth and the related feasibility is analyzed. At the same time, useful exploration has been conducted on the task of segmentation and classification which uses the cascade algorithm of Spectral Clustering and Topic Models. The main research work of the dissertation is as follows: 1. According to the characteristics of motion capture data and the properties of Spectral Clustering algorithms, the advantages and reasons of using Spectral Clustering algorithm on the task of the segmentation and classification of motion capture data are analyzed. This paper proposes an unsupervised method based on Spectral Clustering. First, data pre-processing is performed, and a new distance metric method which suitable for Euler angle data is defined. Spectral Clustering is used to segment the pre-processed motion data preliminarily, and the initial segmentation results are obtained after Statistical Filtering. Based on this rough segmentation, the original motion data is adaptively weighted in combination with the Variance Analysis. The purpose is to calculate the segmented and weighted feature matrix. This matrix is superior to the original motion data in characterizing the motion, and then conduct the spectral clustering again. The re-segmentation is more accurate than the rough segmentation. Among them, the adaptive weighting process based on Variance Analysis also provides a good idea for data initialization for other segmentation classification algorithms. 2. The relationship between motion types and poses or segments in motion capture database likes the relationship between topics and words in a corpus, thus, this paper proposes to use the idea of Topic Model to analyze motion capture data. The similarities and differences between motion data and corpora are compared. This illustrates the possibility of Topic Model ideas applied to the tasks of segmentation and classification and the specific method is proposed. It provides a good ideas for the migration and application of Topic Model. At the same time, the reasons and advantages of choosing the Sparse Topical Model (STM) are also illustrated. A supervised segmentation and classification strategy combining STM and classifiers is introduced. At the same time, with the help of Spectral Clustering-based segmentation algorithm, this paper proposes an unsupervised algorithm Double Spectral Clustering on Sparse Topic Model (DSC-STM). In order to prove the effectiveness of the proposed algorithm, the data in the CMU mocap database was chosen for segmentation and classification experiments. The experimental results demonstrate the effectiveness.