3D Human Pose Estimation Based On Human Body Dynamics

3D human pose estimation is a hot topic in the field of computer vision.3D human pose estimation is the basis for tasks such as human pose recognition,human body tracking,and behavior recognition.It is widely used in the fields of advanced human-computer interaction,intelligent suverliance,et.al.However,it is an ill-posed problem to predict 3D spatial information from monocular images.and the problems such as occlusion and parallax is very challenging.With strong learning capabilities,deep neural networks have been successfully applied in many fields,and have gradually become the optimal choice for 3D human pose estimation.In this work,a phased deep model is proposed to predict the 3D human pose from a monocular image.Specifically,a 2D pose detector is used to obtain the 2D joints,then it is input the 3D estimation model to regress the 3D pose.Under the deep learning framework,human dynamics knowledge is introduced by designing a reasonable network structure and human joint constraints.Experimental results show that the proposed method can relief the prediction errors caused by occlusion,thereby improve the prediction accuracy.The specific content is as follows:(1)Temporal convolutions network structure with multi-stage supervision.Predicting the 3D human pose directly from a 2D image is likely to cause low prediction accuracy,due to problems such as joint occlusion.Therefore,a sequence of images are input to the proposed model,and the detected 2D joints sequence is provided to the 2D to 3D transformation model.In order to further improve the performance,a multi-stage supervision scheme is introduced.(2)3D human pose estimation based on human joint constraints.Because restoring 3D human pose from 2D images is an ill-posed problem.In order to get the specific 3D pose,more knowledge need to be introduced as supervision information to the model.Therefore,human body dynamics is modeled as geometric constraints to the loss function of the depth model.These contraints are learned from the real data of human joint points.Experiments with comparison between knowledge augmented model with baseline demonstrated the effectiveness of the supervised information from the knowledge of human body dynamics.