基于多域融合与特征选择的手势识别研究

Research on gesture recognition based on multi-domain fusion and feature selection

由于手部运动类型增加,准确地对表面肌电(sEMG)信号进行分类需要更多特征。针对特征维数增加导致的特征冗余现象,提出了基于多域融合与特征选择的手势识别方法。从4个通道sEMG信号中提取时域、频域和时频域特征构成特征集。分别采用最小冗余最大相关(mRMR)、基于快速关联的过滤(FCBF)算法、ReliefF、Pearson相关系数进行特征排序,并且利用线性判别分析(LDA)与支持向量机(SVM)对12种精细手势动作进行分类。将4种特征选择与2种分类器的组合构成不同的手势识别模型,SVM-ReliefF手势识别模型分类效果优于其余7种识别模型,在提取30维特征的情况下,分类准确率为96.67%。结果表明:基于多域融合与特征选择的手势识别降维效果显著,且分类效果较好。

As the types of hand movements increase,more features are needed to accurately classify surface electromyogram(sEMG)signals.Aiming at the feature redundancy phenomenon caused by the increase of feature dimension,a gesture recognition method based on multi-domain fusion and feature selection is proposed.The time domain,frequency domain and time-frequency domain features are extracted from the sEMG signals of 4 channels to form a feature set.The minimum redundancy and maximum relevance(mRMR),the fast correlation-based filtering(FCBF)algorithm,the ReliefF,and the Pearson correlation coefficient are used to sort the features,and the linear discriminant analysis(LDA)and support vector machine(SVM)are used to sort 12 kinds of fine gesture motion.The combination of four feature selections and two classifiers constitutes different gesture recognition models.The classification effect of the SVM-ReliefF gesture recognition model is better than the other seven recognition models.The classification accuracy rate is 96.67%when the 30-dimensional features are extracted.The results show that gesture recognition based on multi-domain fusion and feature selection has a significant dimensionality reduction effect and a better classification effect.