Research On Dynamic Gesture Recognition Based On Deep Learning

Human computer interaction has always been regarded as an active research field.At the same time,dynamic gesture is a natural form of communication between people,so it is most suitable for human-computer interaction.There are many problems in dynamic gesture recognition,such as high degree of freedom,appearance difference and time dimension difference.Traditional recognition methods have the problems of poor recognition effect and high equipment requirements,so this paper studies the dynamic gesture recognition method based on deep learning.Through the research and analysis of the existing dynamic gesture recognition methods based on deep learning,some problems are found.Firstly,in terms of deep learning,compared with two-dimensional convolutional neural network,three-dimensional convolutional neural network is more suitable for spatiotemporal feature learning,so as to improve the recognition effect of dynamic gesture.However,3D convolutional neural network has many problems,such as too many training parameters and large model.The optimization work based on three-dimensional convolutional neural network needs to be further explored.Secondly,after the construction of convolutional neural network framework,the number and size of network model parameters obtained by training are usually fixed,which can not be adjusted dynamically according to the actual needs.Finally,in order to improve the accuracy,in the process of network training,the optimization and adjustment of various parameters,the form and content of data input in the network are not enough.In view of the above problems,this paper has carried out the related research and experiment,the main contents are as follows:(1)Compared with two-dimensional convolution,three-dimensional convolution has the problem of large calculation amount and many model parameters.In order to maintain the same accuracy of the model,the standard convolution is replaced by depthwise convolution and pointwise convolution in the three-dimensional convolution process.Designed a 3D depthwise separable convolutional network for gesture recognition.(2)In order to optimize the network model and speed up training and recognition.Based on the 3D depth separable convolutional network,the design ideas of residual connection,group convolution and channel shuffling in the two-dimensional convolutional neural network are used for reference.Combined with the specific requirements of improving the model recognition effect and reducing model parameters in the construction of the 3D gesture recognition network.Designs the 3D depthwise separable convolution residual connection network and the 3D depthwise separable group convolution and channel shuffle network for gesture recognition.(3)In order to dynamically adjust the number and size of network model parameters,the network width adjustment coefficient is introduced.Through the visual analysis of the training loss,training accuracy and model verification accuracy in the process of network model training under different adjustment coefficients,the exploratory work is carried out to balance the recognition results and computational complexity.At the same time,the influence of different input forms and contents on the results of network model recognition is explored through experiments.(4)Using QT visual interface design library,the human computer interaction interface of gesture recognition system is developed,and the method of using the interface is introduced.