基于多流卷积神经网络的动态手势识别

Dynamic Gesture Recognition Based on Multi-Stream Convolutional Neural Network

目的提出一种基于表面肌电信号的多流卷积神经网络动态手势识别方法。方法在表面肌电信号的基础上,引入了加速度以及陀螺仪信号来表征手势的运动特征,并给出了适应输入信号的三分支多流卷积神经网络结构。结果经实自建数据库实验表明,多流卷积神经网络对多类别数据表征的动态手语孤立词有很好的识别效果,不会因数据类别差异而产生的过拟合现象。结论该方法能有效识别多类别数据表征的动态手势,具有一定的创新性。对于其它的多类别数据表征的自然语言信号而言,该结构有一定的参考价值。

Objective To propose a dynamic gesture recognition method based on multi-stream convolutional neural network using surface electromyography signal. Methods On the basis of surface electromyography signal, acceleration and gyroscope signals were introduced to characterize the motion characteristics of gestures, and a three-branch multi-flow convolutional neural network structure adapted to the input signals was presented. Results The experiment of a real self-built database showed that the multi-stream convolutional neural network had a good recognition effect on dynamic sign language isolated words represented by multi-category data and would not generate over-fitting phenomenon due to the difference of data categories. Conclusion This method can effectively identify the dynamic gestures represented by multi-category data and has certain innovation. This structure can be used as a reference for natural language signals represented by other multi-category data.