基于深度卷积神经网络的手势动作识别

Gesture Recognition Based on Deep Convolutional Neural Network

针对传统手势识别方法中人工特征提取信息不完整导致的识别率较低以及识别手势类别较少的问题,基于卷积神经网络(ConvolutionalNeuralNetwork,CNN)的原理,设计了一种深度CNN框架,对多通道的表面肌电信号进行手势动作识别。所应用的表面肌电信号数据来自Ninapro数据库中DB2健康个体数据集,分别识别9种手指动作和49种手势动作(49种手势动作包含9种手指动作),另外40种手势动作是17种基本手势动作和23种手腕动作。对数据集的表面肌电信号数据进行提取均方根值特征,生成12通道的训练集、验证集和测试集。将处理过的表面肌电信号送入到深度CNN中,经过卷积、批次归一化、池化、梯度下降及dropout层处理,仿真测试后,DB2数据集的9种手势动作识别率是99.10%,49种手势动作手势不识别率是64.58%。

In allusion to the problems of low recognition rate and less gesture recognition categories caused by incomplete artificial feature extraction information in traditional gesture recognition methods,a deep convolutional neural network framework is designed based on the principle of Convolutional Neural Network (CNN) to conduct gesture recognition of multi-channel surface electromyography (sEMG) signal. The applied sEMG signal data comes from the DB2 healthy individual data set in Ninapro database, respectively recognizing 9 kinds of finger actions and 49 kinds of gestures,which include 9 kinds of finger actions,and the other 40 kinds of gestures are 17 kinds of basic gestures and 23 kinds of wrist actions. RMS characteristics were extracted from the sEMG signal data in the data set,and 12-channel training set,verification set and test set were generated. And then,the processed sEMG signals are fed into the deep convolutional neural network and processed by convolution,batch normalization,pooling,gradient descent and dropout layer. After the simulation test,the recognition rate of 9 kinds of finger actions in the DB2 data set is 99.10%,and the recognition rate of 49 kinds of gestures is 64.58%.