Font Size: a A A

Research On Human And Camera Pose Estimation In Real Scenes

Posted on:2021-03-30Degree:MasterType:Thesis
Country:ChinaCandidate:T Q SuiFull Text:PDF
GTID:2428330623468515Subject:Engineering
Abstract/Summary:PDF Full Text Request
3D human pose estimation has always been an important area of computer vision research.It is an important prerequisite for the analysis and understanding of human action.By estimating the coordinates of human joint points from images,the human pose is reconstructed in 3D space.There is a wide range of potential needs in virtual reality,video surveillance,sports analysis and other fields.In recent years,with the application of deep learning technology,the 3D human pose estimation technology has improved rapidly.This thesis introduces research works on deep learning based 3D human pose estimation and a improved camera pose estimation algorithm,and finally 3D human pose can be transformed to the world coordinate system based on the estimated camera pose.First,an end-to-end deep network is proposed to jointly learn both the 2D human pose and 3D human pose.Then a GCN-based 3D pose estimation method is proposed.Finally,we use the camera pose estimated from nature image to transform 3D human pose to the world coordinate system The contributions of this thesis are as follows:(1)An end-to-end deep network is proposed to jointly train 2D human pose estimation and 3D human pose estimation.Human body part maps are constructed based on the result of 2D pose estimation and are used to boost 3D human pose estimation.In addition,a incremental refine network is designed to refine 3D pose estimation.The experimental results on Human36M dataset test set indicates that the proposed network can reduce the average joint error by more than 14%compared to the baseline network.(2)A GCN-based 3D pose estimation method is proposed.A multi-scale GCN structure is proposed to improve its representation capacity.Human body part pyramid pooling module is proposed to integrate multi-level features on body parts.The proposed network achieves a result close to the state-of-the-art methods without exploting temporal information.(3)Finally,we propose a improved method for camera pose estimation for sports fields.And the results of camera pose estimation are used to transform the human 3D pose to world coordinate system and restore human 3D pose in real sports scenes.
Keywords/Search Tags:3D human pose estimation, incremental refine network, GCN, human body part pyramid pooling, camera pose estimation
PDF Full Text Request
Related items