基于视觉和肌电信息融合的手势识别方法

Human Gesture Recognition Method Based on Vision and EMG Signal Information

针对人机交互技术对手势识别的可识别种类和识别正确率的问题,提出一种基于视觉方向梯度直方图(HOG)特征和肌电信号(EMG)时域特征融合及支持向量机(SVM)分类器的手势识别方法。利用视觉传感器和智能臂环分别采集手势图像信息和肌电信号,预处理后提取对应的HOG特征和时域特征;采用串行融合的方式将2种特征进行特征级融合;以一对一方式构建组合式SVM多分类器完成手势识别模型的训练和检验,并进行对比分析。结果表明:采用特征级串行融合的36类手势识别模型的总体正确率达到了96%,相较于融合前单一HOG特征和EMG时域特征分别有33%和16%的提升;相较于决策级融合有11%的正确率提升,且计算耗时仅0.274 ms,证明该方法能显著减少特征数据量,有效提高多种类手势识别正确率。

Aiming at the problem of variety and accuracy of gesture recognition in human-computer interaction,a gesture recognition method was proposed based on fusion features of visual histogram of orientation gradient(HOG)and time-domain features of electromyography(EMG),where support vector machine(SVM)was used as a classifier.Visual sensors and smart armbands were used to collect gesture image information and EMG signals respectively,and the corresponding HOG features and time domain features were then extracted after preprocessing.These two features were fused at feature level by employing serial fusion.The combined SVM multiple classifier constructed by one Vs one was used to train and verify the gesture recognition model.Experimental results showed that the overall accuracy of 36 types of gesture recognition models reached 96%by using the proposed fusion features,which is 33%and 16%higher than the single hog feature and EMG time-domain feature before fusion,respectively.Compared with decision-level fusion,the accuracy of the proposed method was increased by 11%,and the calculation time was only 0.274 ms,which could effectively reduce the amount of feature data and significantly improve the accuracy of multiple types of gesture recognition.