Real-time Human Posture Estimation System Based On Deep Learning Study

Human posture estimation is an important research direction in the field of computer vision,and it is also the basis of human motion recognition and analysis and other deeper tasks.Its purpose is to detect the key parts of human body in the image or video,and finally output the parameters of relevant parts.It is widely used in video monitoring,human-computer interaction,digital entertainment and other fields.Human posture estimation is becoming more and more mature,but it is also faced with many challenges.The acquisition of depth information by depth camera can achieve more accurate estimation results of human body posture,and the use of sensors in key parts of human body can also achieve very ideal results,but these are limited by conditions and limited application scenes.People prefer to use ordinary cameras to complete the task of body posture estimation.In this paper,the human body posture estimation is completed with the help of ordinary cameras,and the human body posture real-time control system is built.Firstly,it introduces the theoretical basis of deep learning,which lays a foundation for network design and training.After that,the traditional pose estimation method based on image structure model is compared with the pose estimation method based on deep learning to illustrate the advantages and disadvantages of the two methods.Human pose estimation method to estimate the network using bottom-up,a branch of the network,with the help of convolution posture machine(CPM)algorithm to complete the key points of the human body detection,another branch at the same time,according to the human body key affinity field complete human key points of clustering,combining two branches information,draw 2 d human body key information,after the 2 d human body key information combined with 3 d body posture library,it is concluded that the 3 d human body posture information.The 3D game engine Unity was used to complete the construction of virtual human skeleton and perform skeleton binding.According to the 3D data of key points of human body detected,the spatial rotation information of joints corresponding to the human skeleton was obtained through spatial transformation to control the movement of virtual human skeleton and complete the real-time human posture control system.The experimental results of human posture estimation were cross-verified for different data sets.The 2D and 3D human posture estimation results were compared with the experimental results of relevant scholars using PCK criterion and the average error of each joint(MPJPE),respectively.The experimental results of each scene in the database were significantly improved.A simple filtering algorithm is designed to reduce the fluctuation of virtual human skeleton and realize real-time control of human posture.