Human Pose Estimation Based On Deep Learning

Human pose estimation refers to the process of accurately locating the position of each part of the human body in the image and restoring the posture configuration.Human Pose estimation plays an significant part in the field of human-computer interaction as foundation and it is one of the key technologies for human-computer common intelligence,and it is very important in the field of computer vision.For some general problems and some typical objective problems in human pose estimation,this paper proposes corresponding countermeasures.The main contributions are as follows:(1)A human body pose estimation method based on deformable convolution network is proposed to address the common problems confronted by human pose estimation,such as different scale of human character,different observation angles,articulated human body and all kinds of occlusion.In this method,multi-scale self-inferred human pose estimation depth neural network is as cornerstone.The deformable convolution layers are embedded in the front block to enhance the ability of feature extraction of the whole neural network,enriching the input feature maps of the self-inferred block.This method utilizes the deformable convolution to liberate the plain convolution's constraint which is overly relying on the adjacent pixel relationships.The deformable convolution layers extract the features which is beneficial to solve the common problems.Moreover,the corresponding connection improvement of the self-inference network module is improved,and the utilization effect of the features extracted by the deformation convolutional layer is improved.Experiments have shown that the embedded deformable convolution layers and connection improvement has obtained a more accurate and precise result of the human body pose estimation.(2)A human body pose estimation method based on human body structure integration is proposed.Aiming at there is a issue that the confusions of the left and right corresponding joints,this method uses the part affinity fields and the method of integrating the human body structure to locate the joints' position so that the constraint of the connected joint is strengthened which avoid single joint points being confused by estimation.The penalty loss function is designed to effectively optimize the overall network to enhance the ability of distinguishing the left and right corresponding limbs with similar semantic features.The experiments are proved that the above measures are effective to avoid the risk of confusionbetween the left and right corresponding joints.(3)A human body pose estimation method based on the target individual's attention mechanism is proposed.Aiming at the issue of joint point confusion between target individuals and non-target individuals in multi-person pose estimation,this method proposes to use the individual-attention module to strengthen the network's attention of to the target individual,reduce the interference caused by background information.The using of the selfadversarial network's training mode which is able to normalized heatmaps is added to make the individual-attention module obtain the ability of distinguish the difference between the target individual and the background individual.This ability make the overall generator network a attention-driven network model that accurately obtains the target individual pose estimation results.This method's effectiveness is proven by experiments.