Improved 3D human pose estimation from video based on mixste model

Abstract

3D human pose estimation is an important branch in the field of computer vision. Due to depth blur and self-occlusion, the accuracy of existing 3D human pose estimation methods is low. In order to improve this problem, we propose an improved human pose estimation model based on the Transformer model suitable for processing human skeleton data. We add a pre-training stage to perform the task of recovering 3D skeletons from noisy 2D observations. It helps the model to understand and learn the underlying three-dimensional structure of the human body. By conducting experiments on the public 3D pose estimation data set Human3. 6M and comparing it with currently popular 3D pose estimation methods, it is verified that the above algorithm has a high accuracy.