运动序列的时空结构特征表示模型

运动序列是一种与运动信号相关的多维时间序列,各个维度序列之间具有高耦合性的特点。现有的多维序列表征方法大多基于维度间相互独立的假设或缺乏可解释性，为此，提出一种适用于运动序列的时空结构特征表示模型及其两阶段构造方法。首先，基于空间变化事件的转换方法，将多维时间序列变换成一维事件序列，以保存序列中的空间结构特性。接着，定义了一种时空结构特征的无监督挖掘算法。基于新定义的表示度度量，该算法从事件序列中提取一组具有代表性的低冗余变长事件元组为时空结构特征。在多个人类行为识别数据集上的实验结果表明，与现有多维时间序列表示方法相比，新模型的特征集更具代表性，在运动序列模式识别领域可以有效提升分类精度。

Spatiotemporal structure feature representation model for motion sequences

The motion sequence is a multidimensional time series associated with motion signals, and there are high coupling characteristics between different dimensional sequences. Most of the existing multidimensional sequence representation methods are based on the assumption that the dimensions are independent or lack interpretability. As a result, we propose a spatiotemporal structure feature representation model for motion sequences and its two-stage construction method. First, the multidimensional time series are transformed into one-dimensional event sequences using the transformation method of spatially changing events, preserving the spatial structure characteristics of sequences. Second, an unsupervised mining algorithm for spatiotemporal structure features is defined. A collection of representative event grams of various durations and low redundancy are extracted from the event sequence as spatiotemporal structure features using the newly proposed representation measure. The experimental findings on various datasets for recognizing human behavior demonstrate that the new model’s feature set is more representative than the methods currently used to represent multidimensional time series, and it can considerably increase classification accuracy in the domain of motion sequence pattern recognition.