能量模型数据挖掘的研究

本文通过数据挖掘自身的特点,有效地结合相关算法并基于人体运动捕捉数据,给出这两个问题的有效解决方法。主要工作如下：（1）提出了基于能量模型的算法。相对于现有文献中使用的关节的几何位置,本文提出了的人体能量模型能够有效地降低动作数据的维度,并且能够正确地反映原动作的特征。在此基础上,使用相关系数来表示人体运动过程中各关节之间的相关性,并据此提取出原动作的低维度索引,实验表明该索引能够有效地体现原动作的特征。使用支持向量机结合低维度索引可以有效地讲输入动作划分到一个动作大类中,在此基础上使用基于Keogh下界的线性索引算法可以精确、快速地检索到与输入动作DTW距离最近的候选动作。（2）提出了基于公共子序列距离的数据挖掘算法。相对于现有文献中使用的欧式距离,本文使用的基于最长公共子序列的度量方法能够有效地降低噪声对于挖掘结果的不利影响。使用启发式搜索可以将搜索所需要的时间降低为使用朴素式搜索算法的60%以下,并且随着序列的长度的增加、计算量的增大,前者相对于后者运行时间的百分比有明显的减小趋势,利用这一特性,该算法可以在长序列的主旨模式挖掘中,大规模地减少算法的运行时间。在各长度的候选模式集合中,使用层次化聚类分析可以有效地合并相似度较高的候选模式,以达到合理约简模式、消除相邻重叠模式对结果不利影响的目的。使用最小描述长度原则可以根据模式的长度以及出现频率对候选模式表达整个原序列的能力进行有效地评估,从而达到支持非固定长度主旨模式挖掘的目的。

The Research of the Data Mining Based on Energy model

By taking the human motion capture data as the analysis object,effective algorithm is brought forward on the two subjects respectively andthe experiments verify the availability and high performance.（1）Different from the spatial position based model used in traditional retrieval methods,anovel model based on kinetic energy is proposed and used to describe human motion.Based on the model,the motion coordination is introduced and used toextract the low-dimensional indexing sequences that are remarkable features of original motions.The support vector machine is applied to classify the motion.Finally,the sequential indexing algorithm based on the Keogh lower bound is employed to measure the similarity of query motion and candidate motionsexactly.（2）Against the problem that existing algorithms are apt to be interfered by noise,a motif mining algorithm based on the LCSS distance is introduced.The algorithm has pruned efficiently by using the heuristic strategy based on the distance between subsequences during the search.Then the MDL principleis used to calculate the weights of the unequal-length candidate sequences based on which the motif patterns are selected.