基于卷积神经网络的人体动作识别方法

A CNN-based microwave recognition method for human body actions

针对较小数据集识别时的过拟合和误差传递问题,提出了一种基于卷积神经网络的常见人体动作识别方法.该方法首先利用经典雷达信号处理方法对人体动作回波进行预处理,生成人体动作的时频图像;然后构建卷积神经网络(CNN),并以时频图作为CNN输入数据对网络参数进行训练;最后利用网络公开数据集对所提方法进行了实验验证.实验结果表明,构建的CNN能够准确识别4类不同的人体动作,准确率不低于97%.

Human body actions recognition can not only be used in intelligent home automatic control,but also promote the development of intelligent weapons and equipment.Addressing the over-fitting and error transfer in recognition of small data sets,and the problem of target recognition of linear frequency modulation continuous waves(LFMCW)radar,this paper is based on convolutional neural network(CNN)to propose a microwave recognition method for common human body actions.This method first uses classical radar signal processing method to preprocess the echo of human body actions and to generate time-frequency images of human body actions,and then constructs a CNN,and uses the time-frequency images as input data of the CNN to train the network parameters.Finally,the proposed method is verified experimentally by using the network open data set.The experiment results show that the constructed CNN can accurately recognize four different kinds of human body actions with an accuracy rate of no less than 97%.