Research On Gesture Recognition Based On Deep Learning

Human-computer interaction plays an important role in modern life.How to make people interact with computers more simply,efficiently and naturally is an important research topic and direction.As one of the most natural body languages,gesture recognition has become a key technology in the field of human-computer interaction with the development of computer vision technology.However,in practical application,the influence of surrounding environment and various complex backgrounds brings great challenges to gesture recognition.How to accurately recognize gestures in complex backgrounds has become the focus of current research.At present,the recognition algorithm in complex background mainly uses skin color and depth map to detect gestures.The detection method based on skin color is easily disturbed by illumination,skin-like background and other factors,and its robustness is poor.The depth image needs specific image acquisition equipment,which has serious limitations.Due to the inaccurate extraction of gesture regions from the background,the recognition rate of gesture in complex background is poor.This paper proposes a gesture recognition algorithm HGDR-Net based on in-depth learning.The algorithm first detects gestures based on YOLO,then constructs convolutional neural network(CNN)to recognize gestures.To some extent,it solves the problem of poor recognition effect caused by inaccurate gesture detection,and improves the recognition rate and robustness of gestures in complex background.The main work of this paper is as follows:(1)In the phase of gesture detection,aiming at the difficulty of extracting gesture area in complex background,this paper proposes a gesture detection algorithm based on improved YOLO.This algorithm improves the convolution layer and the convolution core of YOLO model according to the characteristics of gesture.At the same time,it fuses the low-level features and high-level features of YOLO model to provide more location information for final prediction and improve the accuracy of gesture detection and location.(2)In the phase of gesture region recognition,deconvolution operation is applied to visualize the convolution feature map to construct an efficient convolution neural network model.Because the gesture region extracted by detection has a diversity of sizes,and the convolution neural network used for classification requires that the input image size must be consistent.In this paper,spatial pyramid pooling(SPP)is introduced to replace the last one of the convolution network.The layer-pooling layer solves the multi-scale input problem of convolutional neural network,and further improves the performance of gesture recognition.(3)In order to fully train the network model to prevent over-fitting problem,the application of offline data enhancement technology increases the diversity and complexity of training data,but a large number of offline data enhancement makes the training time prolonged.Therefore,in this paper,online data enhancement of training image is carried out during the training process.Finally,two data enhancement methods are combined to greatly improve the network training efficiency and progress.Step enhances the generalization ability of the model.To fully verify the effectiveness of the proposed gesture recognition algorithm,experiments are carried out on NUS-II and Marcel public gesture datasets containing complex background.The experimental results show that the recognition rate reaches 98.65% and99.59% respectively.Finally,the experimental results show that the proposed algorithm can accurately recognize gestures from various complex backgrounds,and has higher recognition accuracy and stronger robustness.