Gesture Recognition Research Based On Deep Learning

The development of computer science has promoted the progress of humancomputer interaction technology.Starting from the interactive method of inputting textual commands in the early days,the human-computer interaction approach has gradually evolved toward the direction of naturalization and visualization.Gestures,as a kind of human-computer interaction,are rich,flexible and intuitive,and are in line with people's daily life and communication habits.In recent years,human-computer interaction based on gesture recognition has gradually attracted the attention of researchers.Gesture recognition methods can be classified into external device based methods and computer vision based methods.Gesture recognition methods based on external devices,such as data gloves,are quite mature and widely used.Computer gesture-based gesture recognition methods are still not mature enough.The traditional gesture recognition methods are generally based on shallow machine learning algorithms such as support vector machines and artificial neural networks.The recognition rate of gestures is difficult to meet the application level requirements,and the traditional algorithm has less room for improvement.In recent years,the rise of deep learning has provided new ideas for gesture recognition.The thesis introduced the relevant theories of deep learning,and mainly explained and analyzed the three commonly used network models of deep learning.Finally,the convolutional neural network that was most suitable for gesture recognition was selected as the focus of this research.Then the main steps of static gesture recognition are explained,and the specific processing method of each step is given.An eight-layer convolutional neural network was initially designed for gesture recognition.Two factors affecting the performance of convolutional neural networks are studied,the choice of gradient descent strategy and the number of neurons in the full connect layer.Finally,the Ada Delta gradient descent method was chosen to achieve the best recognition effect on the gesture test set when there are 500 full-connected layer neurons.Then the research improves the preliminary design of the convolutional neural network gesture recognition method.Firstly,in the preprocessing part of the gesture,elliptical model is used to segment the gesture in the YCr Cb color space.And the processed gesture is further processed to get the two valued gesture sample.The processed gesture samples are more conducive to the extraction of gesture features.Secondly,according to the previous research,the network is improved and a new convolutional neural network structure is obtained.The Inception structure is added to the network to improve the performance of the network.For the given five gestures,the average recognition rate of 98.6% is obtained on the test set.Testing in the actual system can also achieve better recognition results and ensure the real-time performance of the system.Finally,a gesture control browser is designed to apply the gesture recognition method proposed in this paper.The design of the gesture control browser follows the general steps of software design.First of all,browser function is analyzed,and browsers are divided into basic module and gesture module.The basic module is completed through the graphical development environment Qt.And the gesture module defines four kinds of control gestures,which are identified through the convolutional neural network training gesture model proposed in the paper.The entire browser code is written in C++.After testing,when the browser's gesture mode is enabled,the given four control gestures can effectively control the browser.