Research On Gesture Recogntion Technology Based On Convolutional Neural Network

With the rapid development of robot technology,the interaction between machines and humans is more and more frequent,and new requirements for human-computer interaction are put forward.The traditional mouse-and-keyboard-based interaction mode requires people to adapt to the machine,which is unfriendly,difficult and inefficient to be use.The development of technology provides more and more choices for human-computer interaction.The visual-based gesture interaction conforms to human interaction habits,and is more efficient,natural,and friendly.The key technology is gesture recognition technology.This thesis focuses on the deep learning-based gesture recognition technology to carry out the following research work:1.The design and make of gesture datasets for human-computer interaction scenarios:Firstly,the relationship between the recognition accuracy and robustness of the gesture dataset and the model is analyzed.Then the gesture datasets are designed and made to support the training and testing of the subsequent deep convolutional network.2.Deep convolution network parameter selection and module design:Firstly,the thesis points out the feasibility and superiority of applying convolutional network to gesture recognition.Secondly,it analyzes the influence of parameters such as convolution kernel size on parameter quantity,calculation amount,convergence speed and characteristic point receptive field.The thesis uses migration learning method to set initial parameters instead of random initialization parameters to accelerate network convergence and data demand.The thesis introduces the residual network structure to solve the problem that deep network is difficult to train.3.Deep convolution network structure design:The thesis analyzes the characteristics of convolutional layer,pooling layer and batch normalization operation,and analyzes the relationship between network structure and corresponding scene recognition ability.In view of the contradiction between the speed and accuracy of deep convolutional networks,this thesis uses the feature recalibration method to enhance the channel weights that are important for classification and reduce the unimportant channel weights.To solve the problem of small recognition range of deep convolutional networks,a dual-channel convolutional network is designed in this thesis.The network fully utilizes the local feature extraction ability of convolution to enhance the ability of long-distance gesture recognition,and reuses its extracted features for close-range gesture recognition.Implement gesture recognition at different distances.The experimental results show that the proposed gesture recognition method based on deep convolution network can better recognize different kinds of gestures in various complex scenarios at different distances,and provide a feasible way to solve the problem of gesture recognition in different human-computer interaction scenarios.