基于多头自注意力机制和Bi-GRU的人体动作识别算法

Human activity recognition algorithm based on multi-head-self-attention mechanism and Bi-GRU

目前基于惯性传感器的人体动作识别技术具有自主、可靠等优点,但现有人体动作识别算法结构复杂、参数量大、识别精度低。针对以上问题,提出了一种基于多头自注意力机制和双向门控循环单元(Bi-GRU)的人体动作识别算法。该算法首先通过多头自注意力机制无视时间间隔地计算数据之间关联特征,再将关联特征与原始数据拼接,输入到深度Bi-GRU网络中提取顺序特征,最终通过Softmax层识别人体动作。采用YESENSE公司开发的YIS360-V姿态仪搭建了人体动作识别算法验证平台,在此基础上建立训练-测试数据集并进行了实验验证。实验结果表明,所提算法较传统Bi-GRU算法,参数量由40695个减少到18337个,识别准确率由93.36%提升至95.26%。

At present,human activity recognition technology based on inertial sensors has the advantages of autonomy and reliability,however,the existing human activity recognition algorithms have complex structure,large number of parameters and low recognition accuracy.To solve these problems,a human activity recognition algorithm based on multi-head-self-attention mechanism and bidirectional gated recurrent unit(Bi-GRU)is proposed.Firstly,the multi-head self-attention mechanism is used to calculate the correlation features among the data regardless of time interval.Then,the correlation features are spliced with the original data and input into the deep Bi-GRU network to extract the sequential features.Finally,the human activities are recognized through the Softmax layer.The YIS360-V attitude instrument developed by YESENSE is used to build the verification platform of human activity recognition algorithm.On this basis,the training-test datasets are established and the experimental verification is carried out.The experimental results show that compared with the traditional Bi-GRU algorithm,the proposed algorithm reduces the number of parameters from 40,695 to 18,337,and the recognition accuracy increases from 93.36%to 95.26%.