基于改进支持向量机的人手动作模式识别方法

为了提高基于表面肌电信号（sEMG）控制的手部运动康复器对人手多种动作模式的识别率，比较常规支持向量机（SVM）多类分类器的特点，提出改进的决策树支持向量机多类分类方法.该方法引入基于sEMG特征向量的类间距离可分性测度来指导决策树的构建，能够为每个SVM子分类器的训练提供识别率较高的样本划分方案，在提高决策树内部节点分类成功率的同时，简化了分类器结构.通过实验对比可知，新方法在20种手部动作模式的识别训练过程中，单项动作最低识别率较常规决策树方式提高了7.1%，平均识别率达到88.9%，训练速度较一对一支持向量机分类器提高了5.8%.
Abstract: A multi-classification method that employs an improved decision tree support vector machine (DT-SVM) was proposed after detailed comparison of the characteristics of the conventional support vector machine (SVM) multi-classifiers in order to improve the accuracy of hand actions recognition for hand movement rehabilitation device controlled by the surface electromyography (sEMG). The method introduced the measure of distance between classes based on sEMG feature vectors in order to guide the construction of decision tree. The method can provide a sample classification scheme with high recognition rate for the training of each SVM sub-classifier, which can further improve the classification of nodes in the decision tree and simplify the classifier's structure. Experiments for the recognizing process of the 20 kinds of hand motions were conducted. Results showed that the minimum recognition rate of single motion increased by 7.1% compared with the conventional DT-SVM, the average recognition rate reached 88.9%, and the training speed increased by 5.8% compared with the one-versus-one SVM.