基于随机投影的快速稀疏表示人体动作识别方法

A Fast Sparse Representation Classification Method for Human Activity Recognition Based on Random Projection

为有效解决体域网人体行为动作远程识别系统低功耗和快速准确识别的问题,提出一种基于随机投影的快速稀疏表示人体行为动作识别的方法。该方法基于压缩感知随机投影方式压缩数据,获取待测试样本邻近类中较少最近邻训练样本,构建测试样本稀疏表示时的训练样本集,以期达到最优线性重构测试样本;在降低传感器装置功耗和稀疏表示识别算法计算复杂度基础上,捕捉人体行为动作本质特征信息,提高多类别动作识别率。采用国际公开可穿戴传感器动作识别数据库WARD多类别动作数据,验证所提算法的有效性。实验结果表明,当数据压缩率为50%,所提算法能够获得最高平均识别率(92.78%),比传统稀疏表示分类算法获得的动作识别率提高近5%,并显著降低其相应的运行时间,能准确稀疏表示多类别人体行为动作信号,有效降低稀疏表示分类算法的计算法复杂度和运行时间,明显提高多类别动作识别率,为构建快速稀疏表示动作识别提供一个新的思路和方法。

In this paper, a fast sparse representation classification method for human activity recognition based on random projection was proposed, in order to minimize the energy consumption and accurately recognize human activities from wireless body sensor networks-based telemonitoring system of human daily activity. The basic idea of the proposed method is that the random projection way of compressed sensing theory is used to reduce the amount of sampling on sensor nodes within body sensor network, and then the smaller number of nearest neighbor training samples within the neighbor classes of testing sample, which can optimally liner reconstruction testing sample, are obtained to construct the training sample set of the sparse representation of testing sample. Thus, a fast sparse representation classification algorithm with superior performance of generalization can be developed for capturing valuable features of human activity and improving the recognition rate on the basis of the lower energy consumption and computation complexity of algorithm. The multi-class activity data from international open wearable sensor action recognition database WRAD was selected to evaluate the effectiveness of our method. The experimental results showed that, when the data compression rate was 50%, the proposed algorithm could obtain the highest average recognition rate （92.78%）, which was increased by approximately 5% compared with that of the traditional sparse representation classification algorithms. Meanwhile, the operating time of our proposed algorithm was significantly reduced compared with the above traditional methods. We believed that the proposed algorithm could not only effectively reduce the computational complexity and its running time but also significantly enhance the human activity recognition accuracy, providing a new idea and method for developing the fast sparse representation classification algorithm for activity recognition.