Gesture Recognition Method Based On Convolutional Neural Network

Gesture recognition is a research hotspot and a difficult point in the fields of human-computer interaction,intelligent semantic recognition and remote human-machine communication.It is difficult to achieve a high recognition rate in most complicated situations.This paper studies static and dynamic gesture recognition problems,based on the convolutional neural network itself has strong learning ability and the ability to express individual features.For static gesture recognition this paper proposes flattened convolutional neural network and convolutional neural network under complex dense connection.For dynamic gesture recognition the integrating Inception-LSTM concatenated convolution neural network method is proposed.1.For gesture recognition problem in simple background,this paper uses Kinect depth camera to collect experimental data samples of different gestures,under the Tensor Flow framework with flattened convolution module proposing a gesture recognition algorithm-FD-CNN network.Firstly,preprocessing the dataset,and then the data is input into the FD-CNN network for training.Compared with the deformable convolutional neural network,it is found that the recognition rate of this paper method is high.2.The FD-CNN network has a good recognition effect in a simple background image with significant features,but it has poor effect on static gesture recognition under complex background.Therefore,for complex background gestures this paper proposes a convolutional neural network based on complex dense connections and uses SGD(Stochastic gradient descent)and Adam(Adaptive Moment Estimation)network optimization strategies to optimize the network.Experiments on the Thomas Moeslund and Kinect Leap gesture databases respectively showed that the recognition rate is higher in RGB images and depth images.3.For dynamic gesture recognition problem,this paper fusion constructs the Inception-CNN network and LSTM network structure.Firstly,encoding the input video sequence,and then using the tower parallel Inception-CNN network extracts the feature.The extracted information is merged into a dynamic spatiotemporal sequence and then classified and identified using LSTM network.Verification in three different dynamic gesture datasets of Cambridge-Gesture,VIVA and Sheffield Kinect Gesture Dataset(SKIG)shows that the fusion Inception-LSTM cascade network has high recognition rate.Comparing with support vector machine(SVM)and direction gradient histogram(HOG)and convolutional neural networks,the method achieves better recognition results,which proves the effectiveness and robustness of the proposed method.