摘要
为了提高人体连续运动检测识别的准确率和效率,提出一种基于卷积神经网络(CNN)的识别检测方法.首先,通过3轴加速度传感器采集运动加速度数据,经过滤波、滑动时间窗口和快速傅里叶变换(FFT)获得加速度频谱以及运动模式识别的特征矩阵.通过对多种结构与参数的网络模型的测试、分析与研究,选择双层33卷积核卷积层加双层256 ×64神经元稠密层的CNN结构,分别对站立、步行、上楼、下楼和跑步5种典型运动和跌倒、倒卧2种异常运动进行识别.结果表明,本文方法对跌倒、倒卧等7种运动的识别准确率大于95％,是人体连续运动检测识别的一种有效的方法;同时本文方法对算力需求较低,可应用于低功耗移动平台.
Abstract
To improve accuracy and efficiency of the continuous human motion detection and recognition,a recognition and detection method based on convolutional neural network(CNN)is proposed.Firstly,motion acceleration data are collected through a 3-axis acceleration sensor,and acceleration spectrum and the characteristic matrix of motion pattern recognition are obtained through filtering,sliding time window and fast Fourier transform(FFT).Through testing,analysis and research on network models with multiple structures and parameters,a CNN structure with a double-layer 33 convolution kernel convolutional layer and a double-layer 256 × 64 neuron dense layer is selected to recognize 5 typical exercises of stand-up,walk-up,upstairs,downstairs and running and 2 abnormal exercises of falling and lying down,respectively.The results show that the recognition accuracy of 7 kinds of movements,such as falling and lying,is greater than 95％,and it is an effective method for the detection and recognition of human continuous motion.At the same time,the method has a low demand for computing power and can be deployed in low power consumption on mobile platforms.
基金项目
国家电网公司科技项目(5600-201926164A-0-0-00)
维普
万方数据