基于卷积神经网络的手势识别算法  

Convolution Neural Networks for Gesture Recognition

作　　者：朱雯文 叶西宁[1] ZHU Wen-wen;YE Xi-ning(School of Information Science and Engineering,East China University of Science and Technology, Shanghai 200237,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237

出　　处：《华东理工大学学报（自然科学版）》2018年第2期260-269,共10页Journal of East China University of Science and Technology

基　　金：国家自然科学基金(60974066)

摘　　要：手势识别是人机交互、智能假肢、医疗康复等领域的研究热点。为了满足手势识别实时性和准确性的需求,本文以成本较小的加速度信号作为数据,在对LeNet-5卷积神经网络进行分析的基础上,提出了一种适合加速度信号的LeNet-A网络。该网络针对基于加速度的手势分类特有的复杂性,增加Dropout层,改变卷积核大小、卷积核数量、激活函数以及分类器。在Ninapro数据集上的实验结果表明,该网络在正常受试者和截肢者的识别率上均表现出很大的优势,平均精度分别为90.37%和79.99%,比目前最佳分类器提升了12%和31%左右。该网络还具有较好的实时性和抗噪性。Gesture recognition is a hot topic in the fields of human-computer interaction,prosthesis control,rehabilitation,and so on.To satisfy the requirement of real-time and accuracy on gesture recognition,this paper makes the analysis on LeNet-5 and proposes the LeNet-A network suitable for acceleration signal.In the proposed LeNet-A,the low-cost acceleration signal is taken as the input of network,which are six layers,including two convolution layers,two pooling layers,one fully connected layer and one output layer.Especially,by considering the specific complexity of gesture recognition based on acceleration signal,we introduce the Dropout layer and change the size of convolution kernel and the number of convolution kernel.Besides,the activation function is replaced by Relu function and the classifier is replaced by Softmax classifier.This paper chooses Ninapro database as the dataset,which is an available public resource for research on gesture recognition and has up to more than 50 gestures including intact subject and the amputated.Before making classification,these signals in this dataset are dealt with by low-pass filtering,down-sampling,data balance and training/testing set segmentation.The simulation results show that the proposed network has great advantage in the gesture recognition both in the intact subject and the amputated and attain the average accuracy of 90.37%and 79.99%,respectively,which increase by 12%and 31%,compared with the best result in the exiting literature.The test on one subject with and without rest gesture shows that the accuracy without rest gesture is 93.60%,increasing by 3.1%compared with the accuracy with rest gesture.It is observed that the rest gesture has a relatively large influence on the whole gesture classification accuracy.The latter experiment also illustrates that this proposed network not only improves the gesture recognition accuracy distinctly with characteristics of good real-time,but also has quite strong robustness to noise.These further verify that the proposed method

关 键 词：加速度 卷积神经网络 手势识别 抗噪性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]