一种用于鉴别体域网动作模式的近邻快速鲁棒协作表示分类算法

A Novel Neighboring Collaborative Representation Classification Algorithm for Recognizing WBAN-Based Activity Pattern

提出一种近邻快速鲁棒协作表示体域网动作模式分类算法,旨在基于体域网多传感动作模式数据结构内在相似性,利用最近邻原则,寻找与测试动作样本密切相关的少量近邻类别和近邻训练样本,重新构造训练样本集,然后基于新训练样本集构建快速鲁棒协作表示动作分类模型,通过扩展拉格朗日乘数算法求解待定测试样本协作表示系数和表示残差,定义判定测试样本所属类别规则,有效提高分类性能。采用公开的美国加州伯克利大学多传感动作模式数据库验证所提算法有效性。结果表明,所提算法能够从体域网多传感数据中获得更多与动作模式密切相关的协调性和相关性,动作模式识别率提高2%,运行时间仅需6.5 s,分类性能明显优于稀疏表示动作模式分类性能,有望为临床鉴别人体动作模式提供一个新的技术解决方案。

This paper proposed a novel neighboring robust collaborative representation algorithm for wireless body area networks（WBANs）-based activity classification. Based on the similarity of multi-sensor action data structure, our proposed technique found out a few neighboring classes and samples associated with test sample according to the nearest neighbor principle. This allowed to construct the new training set for collaboratively representing action patterns. And then, the augmented lagrange multiplier algorithm was adopted to solve the representation coefficients and representation residuals of test sample, in order to significantly improve the classification performance. The multi-sensor action data are selected from an open wearable action recognition database（WARD） of University of California, Berkeley, in order to validate the effectiveness of our proposed technique. The results showed that our proposed method could capture more valuable correlation information associated with human action. The best accuracy was increased by 2% and the running time only spends 6.5 s, which suggested that our proposed technique was obviously superior to the sparse representation-based action classification algorithms. It is very helpful to offer a new powerful tool for recognizing action pattern in clinical application.