3D Human Pose Estimation Based On Monocular View

With the development of image sensor technology,image-based human motion analysis has become a hot topic in the field of computer vision and data mining.The 3D human pose estimation is an important premise for the analysis and understanding of the motion,which estimates the coordinates of the key points of the human body,the rotation angle and other parameters from the image,and helps with the reconstruction of the human pose in three-dimension.The 3D human pose estimation has broad application prospects in the fields of motion analysis,virtual reality,film production and so on.This dissertation mainly studies the three-dimensional human pose estimation problem based on monocular view.This problem can be divided into two steps: extracting the two-dimensional pose of the human body from the image and estimating the three-dimensional pose by using the two-dimensional position coordinates.Considering that the existing two-dimensional human body pose estimation scheme is relatively mature,this dissertation specifically aims to reconstruct the depth information of human pose under the condition of knowing twodimensional human body position.The human pose has a high continuity and correlation in the time dimension.Therefore,based on the time series data of the two-dimensional pose of the human body,the three-dimensional pose estimation model is realized by using the Sequence to Sequence framework and the LSTM unit.Then the model is improved by introducing the attention mechanism and the acceleration data of the human joints from the inertial measurement unit.The average error on the TotalCapture dataset is 39.03 mm,and the prediction accuracy was improved by 6.3% after the improvement.The experiment proves that it can effectively reduce the prediction error of the model and improve the prediction accuracy of the model on the distal node of the human body.The robustness of neural networks depends heavily on the scale and coverage of the training dataset.However,the existing 3D annotation datasets have relatively simple scenarios and actions.Therefore,this dissertation studies the unsupervised learning of 3D human pose estimation.Based on the Generative Adversarial Network,with the mutual game between the generator and the discriminator,and the connection between the three-dimensional pose and the two-dimensional imaging result,the unsupervised model using only the two-dimensional annotation datasets is realized by combining the prediction results of 3D human pose with the random projection of the elevation angle and the plane rotation.At the same time,in order to evaluate the effect of the model more intuitively,this dissertation combines the existing 2D pose estimation algorithm to complete the model visualization of the three-dimensional human pose estimation based on monocular image.