利用表面肌电进行手势自动识别

Automatic Gesture Recognition with Surface Electromyography Signal

针对手势自动识别研究中提高正确率和降低训练时间两者需要同时兼顾的问题,提出了一种基于Fisher Score(FS)特征降维方法与机器学习相结合的新的手势识别模型。提取4通道表面肌电信号的时域、频域、时-频域和非线性特征,构成特征集;采用FS方法和主成分分析(PCA)方法分别进行特征降维,采用线性判别分析(LDA)和支持向量机(SVM)分别作为分类器;通过两种特征降维方法与两种分类器的不同组合构建不同的手势识别模型,并对分类模型的性能进行对比研究。实验结果表明,特征降维方法与分类器的组合能显著提高分类器的正确率、降低训练时间。与PCA方法相比,FS方法是一种实现简便、效果理想的特征降维方法:与SVM组合的分类模型获得最高分类正确率99.92%;与LDA组合的分类模型不仅获得99.24%的分类正确率,而且花费最短的训练时间1.44ms,该模型可为手势的实时自动识别提供理想的方法和途径。

For the aim of improving accuracy and reducing training time in automatic gesture recognition,a new gesture recognition model based on Fisher Score(FS)feature reduction combined with machine learning method is proposed.Firstly,the feature set is extracted from four-channel surface electromyography signals,involving the features of time domain,frequency domain,time-frequency domain and nonlinear dynamics.Then,FS and principal component analysis(PCA)are used for feature reduction respectively,and linear discriminant analysis(LDA)and support vector machine(SVM)are adopted as classifiers.Finally,different gesture recognition models are constructed by the two classifiers with and without two feature reduction methods,and their classification performance is compared.Experimental results demonstrate that the feature reduction method can help classifier improve accuracy and reduce training time significantly.Furthermore,compared with PCA,FS is a simple and effective feature reduction method:the combination of SVM and FS can achieve the highest classification accuracy of 99.92%;the combination of LDA and FS obtains the accuracy of 99.24%and takes the shortest training time of 1.44 ms,which suggests an ideal method for real-time automatic gesture recognition.