Research On Human Posture Recognition Technology Based On OpenPose And Human-Computer Interaction

With the rapid development of deep learning and computer Vision,it has gradually entered the public life from the theory.With the rapid development of deep learning and computer Vision,human posture recognition technology has gradually occupied an important position in the fields of human-computer interaction,medical rehabilitation,monitoring and security and smart home.The specific process of human posture recognition technology is to input data such as images or Videos,detect and recognize key points of the human body on two-dimensional image data,and obtain human posture information through key points to understand human behavior.Although the human posture recognition technology develops rapidly,the network model is gradually enlarged,the number of network layers is gradually deepened,and the identification accuracy of human key points in complex environment is gradually improved.However,in practical application,a single network model cannot make full use of different features of the image,and it is easy to cause problems such as false detection and missing detection caused by factors such as occlusion generated by complex environment.In addition,the deepening of the number of network layers will increase the requirements of the model on the hardware,and reduce the identification rate.Meanwhile,the huge number of parameters makes it easy to produce problems such as gradient disappearance or gradient explosion during training,which increases the difficulty of training.Making time-sensitive humancomputer interaction and other projects difficult.In the previous research work,usually focus on the improvement of a single module of the network but ignore the improvement and innovation of the overall network scheme.After fully studying RGB color image,RGG-D depth image,Open Pose,Goog Le Net and other networks,due to the different data types of color image and depth image,a two-channel convolutional neural network with excellent performance for RGB image data and depth image data as input is used.Namely,the two-branch convolutional neural network composed on the basis of Goog Le Net and NIN structure is used for human posture recognition.Meanwhile,the Inception structure and NIN structure are improved so that the two branch networks can fully extract the semantic features of RGB data and depth data,and conduct feature fusion experiments on the input stage.Deconvolution is determined to be used for the intermediate stage feature fusion of the input domain,so as to achieve the effect of complementary data extraction by using the optimal network,and complete the human body pose recognition.The NAdam optimizer was used to optimize the neural network.Experiments show that the improved Inception structure and NIN structure of the dual channel convolutional neural network not only improve the recognition accuracy,but also greatly optimize the number of parameters,performance is significantly better than the original network and other mainstream neural networks.