Research On Gesture Interaction Method Based On 3D Pose Estimation FROM Single RGB Image

Gesture interaction is an interactive method in which the computer executes the corresponding instructions and actions by acquiring the data and information expressed by people ’ s hands.Gesture interaction has application value in various fields,and because the hand is the most natural in people ’ s daily use,the hand for virtual scene interaction has very important significance.In recent years,in order to improve the interactive experience,interactive method of hand recognition has been a hot topic.There are still some common problems in hand interaction research.Firstly,most gesture interaction methods require hardware devices with fixed conditions,which will lead to increased interaction costs and increase the difficulty of users.Secondly,many research methods that do not need hardware devices will lose the original threedimensional information of the hand when interacting,resulting in the lack of realism in hand interaction.In addition,with the increase of the complexity of the processing algorithm,the processing speed of hand interaction is also getting slower and slower,resulting in many interactive researches empty high accuracy but cannot be widely used.To solve the above problems,firstly,in order to reduce the demand for hardware equipment,the single RGB image is processed,so that the interactive operation can be completed by using the basic monocular camera.Secondly,the three-dimensional gesture estimation method is used to estimate the lost three-dimensional information of the hand from the single RGB image,and a gesture semantic classification method is proposed to classify the gesture semantics.Finally,the two methods are integrated to complete the final gesture interaction.In addition,the deep separation convolution,residual network and feature superposition are used to improve the speed and accuracy of 3D gesture estimation method.The decision tree,random forest and limit tree are used to combine the voting classifier to realize fast gesture semantic classification,so as to ensure the gesture interaction with both speed and accuracy.The three-dimensional gesture estimation network is trained on two public datasets to ensure the generalization ability.The verification results show that the hand estimation method can consider the processing speed and accuracy.The gesture classifier is trained on the annotated public dataset,and the results show that the speed and accuracy are considerable.In order to verify the feasibility of the gesture interaction method,a virtual interaction test platform is finally created.Through the test,the feasibility of the proposed interaction method is proved,and the operation effect is good.