Human Pose Estimation Based On Convolutional Neural Network

Human pose estimation is a technique for obtaining the position and orientation of human joints from an image,that is,outputting the joints of the human body and the connections between the joints.Human pose estimation is one of the hot research directions in computer vision,and it has a wide range of application requirements in the fields of medical treatment,security,and virtual reality.This paper focuses on the research of human pose estimation based on convolutional neural network.After the in-depth study of existing human pose estimation methods,an error-correcting matching model with associated joint pair constraints is proposed and combined with the Convolutional Pose Machines(CPMs)for single-person pose estimation.The experimental results prove that the proposed method can significantly improve the accuracy of joint positioning,and correctly connect the joints that correspond to physiological characteristics.Furthermore,Multi-Channel Simplified Convolutional Pose Machine(MULC-SCPM)network is proposed and combined with regional positioning to achieve multi-person pose estimation.These experimental results prove that based on the localization,the introduction of the MULC-SCPM network effectively improves the accuracy of joint connections under occlusion conditions,and achieves the effect of improving the accuracy of multi-person pose estimation.The main work of this paper focuses on the following three aspects:First and foremost,learn and analyse three types of human pose estimation methods,which are methods based on overall features,models and deep learning,and implemented several representative human pose estimation methods based on deep learning.The research focuses on the deep learning human pose estimation method based on heatmap regression,and illustrates that high-quality human pose estimation results are of great significance in practice.Additionally,propose an error correction matching model with associated joint pair constraints,and combine the proposed model with Convolutional Pose Machines(CPMs)network to estimate human pose.The construction of the model takes into account the human body structure and physiological characteristics.For instance,the human body can only have one correct right or left associated joint pair connection,and the distance between the left and right hip joints must be greater than the distance between two types of joint on the same knee range.The proposed model can correct the mis-calibration of joints,and correctly connect the joints that match the physiological characteristics.The experimental results demonstrate that the Convolutional Pose Machines(CPMs)with associated joint pair constraints show good results in single-person pose estimation,especially it can effectively solve the problems that joints are repeatedly counted and reversed in wrong order.Last but not least,present a multi-person pose estimation method combining region localization with MULC-SCPM network.This method firstly introduces region positioning to achieve single-person region segmentation;then pose estimation is performed based on Multi-Channel Simplified Convolutional Pose Machine(MULC-SCPM)network that combines joint position loss,connection distance loss,and direction loss.Region positioning can effectively determine the position of the human body.The MULC-SCPM network can effectively return the position of human joints and the connection relationship between joints.The experimental results indicate that the regional positioning combined with the MULC-SCPM network performs well,especially it can effectively solve the problem of joint connection error when the target is partially occluded.The experimental results demonstrate that the methods can improve the accuracy of single-person and multi-person pose estimation.