3D Human Motion Reconstruction From Monocular Video

It is hoped that computers can understand the real world as well as humans,and human body movement,as one of the most common visual contents in the real world,is a classic research object of machine vision.Motion reconstruction,as the basic algorithm for high-level application,represents human motion by recovering the coordinate of joint points and rotation Angle from images and videos,which has a broad application prospect in robot perception and simulation,behavior analysis,virtual reality and other fields.Under the promotion of artificial intelligence,two-dimensional motion reconstruction has been used by the industry,but the world is three-dimensional,two-dimensional human body expression cannot accurately restore the posture of the human body,in order to better understand human motion,more challenging three-dimensional motion reconstruction has become a research hotspot in recent years.The existing 3d motion reconstruction methods are mainly divided into two categories:the method based on data mapping of different dimensions and the method of parameter prediction of human model.However,these methods have the following problems:difficult to take advantage of the hierarchical connections between human joints;There is no consistency of human model on time sequence,and only coordinate information is transferred between different prediction frames.It is difficult to label the 3d human joint coordinates,and the over-fitting generated by training on the limited data set makes the performance of the algorithm in the real scene not ideal.To solve these problems,this paper proposes a 3d human motion reconstruction method based on monocular video and predicts the fixed human skeleton and the corresponding joint rotation,global coordinates and footstep contact signals based on the dual-path neural network structure designed.By simulating the forward kinematics in the neural network,a single skeleton can be inserted into the whole time series prediction,which greatly reduces the search range of joint coordinates and ensures that the prediction results conform to the kinematics knowledge.The antagonistic learning used in joint angular velocity also guarantees the authenticity of the generated results.In this paper,the effect of the algorithm has been evaluated on several public data sets.The experimental results show that the proposed method can recover the most accurate human skeleton and the reconstruction results are very accurate.Compared with other methods,it is found that this method greatly improves the accuracy,authenticity and robustness of motion reconstruction.