Abstract: A novel and appropriate algorithm is the key of human activity pattern recognition system. This paper proposes a human activity pattern recognition method utilizing accelerometer data, which adopts a multi-graph embedding representation algorithm for feature dimensionality reduction and a nearest neighbor classifier for pattern classification. Firstly, the proposed method divides the original feature space into several disjoint subsets by clustering so as to generate multiple graphs. Then, the method generates embedding co-ordinates with multidimensional scaling and finds the linear combination of the embedding co-ordinates to represent the original feature space. Finally, the method adopts a nearest neighbor classifier to classify the patterns. The proposed method is novel and simple, and can explore the potential structure of the original feature space with the smallest information loss so that the feature selection is more stable. The experimental results show that the proposed method has a higher accuracy than other representative methods, and can better recognize human activities.

Key words: accelerometer, human activity pattern recognition, multi-graph, multidimensional scaling, graph embedding

摘要： 新颖和恰当的算法是人体运动模式识别系统的关键。在获取加速度传感器信号的基础上，提出了一种人体运动模式识别算法，其中多图嵌入表示用于特征降维，最近邻用于模式分类。该算法通过特征分组对原始特征空间进行多个独立子集的划分，并生成图；通过多维尺度分析法在每个子图上生成新的嵌入坐标，并找到这些嵌入坐标的线性组合来表示原始特征空间；最后通过最近邻分类器进行模式分类。该算法新颖、简单，能在最小信息丢失的基础上挖掘原始特征空间的潜在结构，提高特征选择的稳定性。实验结果表明，同其他代表性算法相比，该算法准确度高，能更好地区分人体运动。

关键词: 加速度传感器, 人体运动模式识别, 多图, 多维标度法, 图嵌入