Recognition Of Upper Limb Posture Based On Deep Convolution Neural Networks

Human pose recognition is a theme in computer technology and language technology.It is aim to get human posture information through computer algorithm and understand meaning.It refers to the process of identifying human joints by capturing human motion data.In recent years,with the promotion of computer hardware upgrading and software algorithm optimization,it has made great progress,and is more and more widely applied in the fields of auxiliary clinical diagnosis,rehabilitation engineering,human motion analysis and so on,showing great application value and there.However,human gesture recognition based on RGB images is vulnerable to interference from light,environmental noise,complex background and human movement.It is difficult for classical methods to take account of the contradiction between model recognition accuracy and generalization ability.Meanwhile,there is still a gap between application requirements from real-time processing.Therefore,in view of the natural light human attitude recognition problem,we must overcome the decline of gesture recognition rate and even fail to recognize it when meeting the real-time requirement of the system,which is the focus of this paper.Thanks to the attention of deep learning achievements in the field of machine vision and pattern recognition,this paper deep convolutional neural network based on the theory of series of problems around the natural light image attitude recognition,designed to drive data as a machine learning method to overcome the problems in the traditional method.Therefore,in view of the natural light human attitude recognition problem,we must overcome the decline of gesture recognition rate and even fail to recognize it when meeting the real-time requirement of the system,which is the focus of this paper.Thanks to the attention of deep learning achievements in the field of machine vision and pattern recognition,this paper deep convolutional neural network based on the theory of series of problems around the natural light image attitude recognition,designed to drive data as a machine learning method to overcome the problems in the traditional method.The main work of this article is as follows:1.A fast and effective machine learning denoising algorithm is proposed for image preprocessing,which seriously affects image quality and directly leads to image noise reduction.The total convolution neural network(full convolution neural network)is introduced to perform noise mask pixel value regression operation to generate an estimated noise template.The image denoising purpose is achieved by weakening the noise influence in the noise image.Compared with the same type of machine learning method,this method is faster than the same type of machine learning method.Compared with the classical algorithm,the algorithm greatly improves the real-time performance of the algorithm and shows better denoising effect.2.The core problem of RGB image human pose recognition is studied.The pose recognition problem is defined as the coordinate point regression problem.Aiming at the 14 joints of human body,the C-pose model is designed to realize the coordinate location in the image.The C-pose algorithm transforms the traditional classification network into an end to end model to solve the regression problem,and produces the corresponding coordinates of the joint point directly in the final full connection layer.At the same time,cascade training learning mode is introduced to optimize the visual data range of the model and improve the recognition accuracy.Compared with some mainstream algorithms under the classical metric standard PCP,the highest average recognition accuracy is increased by about 11.9%.Compared with the classical method,the model has higher recognition accuracy and more generalization ability,and it can meet the real-time requirement of the system with GPU operation equipment.3.In this paper,we study the problem of human posture semantic understanding.We use C-pose gesture recognition results to construct human joint tree diagram.We classify 7 common gesture classifications through classical classification network,and indirectly achieve simple human posture semantic understanding.The three basic attitudes of standing,walking and sitting are above 70% recognition rates,which meet the requirements of real time,and have practical application feasibility.