Research On Gesture Recognition Based On Autonomous Learning

In recent years,based on the continuous development and progress of science and technology,human-computer interaction has also further developed to the direction of intelligence.As one of the body language,gesture can express a lot of information intuitively and effectively.It is also convenient for some special groups such as deaf and mute people to use it.Therefore,the research value of gesture recognition is high.This paper analyzes and discusses the related technologies of gesture recognition based on machine learning.The main work contents and achievements are as follows:(1)After capturing the gesture,how to accurately extract the gesture area from the image.In order to solve the problem that gesture is easily affected by light,this paper studies the segmentation effect of several common color spaces on gesture under the same skin color.Combined with the actual scene,the experimental analysis shows that the component of YCb Cr color space is more stable,and the segmentation effect is better than HSV and RGB.Therefore,YCb Cr color space is selected to establish skin color model;however,if there is a class,the color space of YCb Cr is selected to establish skin color model In the presence of skin color objects,gestures can not be segmented well,so a rule of excluding non gesture regions is proposed.Finally,the minimum bound rectangle(MBR)and minimum area bound rectangle(MABR)of hand region are studied to enhance the basis of gesture recognition model.(2)Aiming at the gesture recognition model RF-Net,we first add BN layer and1*1 convolution layer on the basis of Alex Net,and adopt dynamic learning rate to get Alex Net?I model,and then keep the network parameters of the model unchanged(except for the full connection layer),and extract the features of the segmented gesture image.Because random forest has the characteristics of strong over fitting resistance and high recognition rate,and its calculation strategy can effectively reduce the identification time,which is the decision framework of RF-Net model.The model combines the characteristics of convolution architecture and decision architecture.Through training and testing in Marcel data set and HGHA data set,the accuracy and recognition rate are improved,which shows that the segmentation algorithm and the recognition algorithm based on RF-Net model are practical and effective.(3)For the application in the actual scene,a gesture recognition system is developed.Firstly,the software and hardware environment of the gesture recognitioncontrol system are described,and the gesture acquisition module and gesture recognition module are introduced respectively.In the aspect of identification module,the robustness of the system is verified.