一种基于双流融合3D卷积神经网络的动态头势识别方法

A Dynamic Head Gesture Recognition Method Based on 3D Convolu⁃tional Two⁃Stream Network Fusion

目前基于视觉的动态头势识别算法泛化能力弱、识别率低,头戴式传感器的方法经济性、便携性差.针对以上问题,提出了一种无需头戴设备的动态头势识别算法.这种基于双流融合3D卷积神经网络的方法用头部动作生成稠密光流,并将原始数据和光流数据并行输入构建的动作特征提取器,最后进行特征融合.结果表明所提算法比人工特征提取方法和C3D模型有更高的准确率、更好的泛化能力,在无需头戴传感器的情况下有近似头戴式传感器的识别率.

Present vision based on dynamic head gesture recognition algorithms usually have disadvantages in gener⁃alization and recognition rate,and head⁃mounted sensors are expensive and inconvenient.In view of the above problems,a dynamic head gesture recognition algorithm without head⁃mounted sensors is proposed.Using this method based on two⁃stream 3DCNN(3D Convolutional Neural Network),the dense optical flow is generated by head movements,the origi⁃nal data and dense optical flow are put into the motion feature extractor in parallel,and finally,features are fused.Experi⁃mental results show that the proposed algorithm has higher recognition accuracy and better generalization than the artificial feature extraction and C3D(Convolutional 3D)methods,and its recognition rate is as good as those head mounted sensors.