Bone-based Method For Balanced Singer-person Human Pose Estimation

With the continuous development of society and the improvement of people's living standard,people begin to pay more attention to the importance of human body attributes in the progress of science and technology.The development of relevant application technology and analysis of massive image and video data by using human posture extraction technology has become the mainstream trend of Internet development.As the basis of human posture extraction technology,pose estimation has become the basic technology in many fields.Compared with other tasks,the joint points of human posture are smaller in scale,and have strong interference to occlusion and fuzzy.Therefore,the algorithm has high requirements for accuracy and robustness in the design process.Currently,most pose estimation algorithms are based on neural network algorithm,and image and video features are extracted by complex neural network,and then the joint points are extracted according to the characteristics of the point.However,because the network does not deeply dig the internal relationship between human body and joint points in the design process,it is easy to achieve bottleneck by simply relying on the improvement of network complexity,which makes the accuracy rate difficult to continue to improve,and the algorithm complexity is too high to meet the actual use needs.This thesis proposes a single person pose estimation method based on human bone for human body in monocular image.This method can improve the accuracy of single person posture greatly.The main work of this thesis includes the following two aspects:(1)Most of the pose estimation algorithms use the image information around the joint points to classify and locate the joint points,which leads to the imperfect use of human information.When the image semantic information of human joint points is seriously lost,it is difficult to locate them.To solve this problem,this thesis proposes a single person pose estimation network based on human bone.This network uses the information around the joint points to supervise the network,at the same time,it models the bone and the overall architecture of the human body,and integrates with the features of the joint points to enhance the representation ability of the network on the human body structure,which is helpful to the further prediction and accuracy optimization of the joint points.At the same time,this thesis also puts forward the guidance idea of difficult and easy joint points,through the simple joint points to guide the position of difficult joint points,improve the prediction accuracy of difficult joint points.Experiments show that this strategy can make the human joint information be fully completed,and the detection accuracy of human joint points is greatly improved.(2)Aiming at the problem of large deviation between the loss function and the learning target in the single person attitude estimation network,this thesis proposes a balanced loss function of pose estimation based on classification.The thermal graph of human joint points and the bone graph of human body contain abundant human spatial information and semantic information.How to optimize them with reasonable loss function is an important step to predict the joint points.Most of the current human pose estimation algorithms regard the task as a regression problem,and use the mean square error loss function to complete the point-to-point restoration of the target heatmap.In this thesis,the task of human pose estimation is transformed into a classification problem,and the difficulty of joint detection is further modeled,so that the final heatmap prediction results focus on the area around the joint.Experiments show that this method can improve the training efficiency of the model and make the network locate the joint points more accurately.To sum up,the balanced pose estimation network based on human bone proposed in this thesis can achieve accurate pose estimation results in monocular images of human body,which is higher than the accuracy of existing pose estimation models,and has certain theoretical research value and practical application value.