3D Human Pose Estimation Based On Domain Adaptation Key Technology And Prototype Implementation

3D Human Pose Estimation is an important research topic in computer vision,and its goal is to estimate the position of human key points from images or video sequences.The 3D human pose estimation algorithm is divided into single-stage and two-stage.The core idea of the two-stage method is first to estimate the 2D human pose and then upgrade the 2D human pose to 3D.Compared with the single-stage algorithm,its network structure is simpler,and the prediction accuracy is higher.Therefore,two-stage 3D human poses estimation methods have become the focus of researchers.However,lifting 2D poses to 3D faces the problems of 2D-3D pose ambiguity and insufficient labeled datasets.The research in this thesis introduces domain adaptive technology into 3D human pose estimation to solve the above two problems.The main work content of this thesis is as follows:(1)A 2D-3D pose lifting network based on multi-grain feature alignment is proposed.Aiming at the problem of 2D-3D pose ambiguity in the process of 2D pose lifting to 3D,this thesis proposes a multi-granularity feature-aligned 2D to 3D pose lifting network.2D human pose and 3D human pose are considered as two domains,first extracting multigrained 3D human prior knowledge from the 3D human pose and then transferring the extracted prior knowledge to 2D-3D pose lifting network by domain adaptive techniques.The proposed method strengthens the 3D human pose generator’s understanding of the human pose.Compared with other methods,the experimental results show that the proposed method has smaller prediction errors and higher accuracy.(2)An unsupervised domain-adaptive 2D-3D pose lifting network is proposed.Aiming at the problem of insufficient labeled datasets in the process of lifting 2D poses to 3D,this thesis proposes an unsupervised domain-adaptive 2D-3D pose lifting network.The dataset with 3D pose labeling is regarded as the source domain,and the dataset without 3D pose labeling is regarded as the target domain.The network improves prediction accuracy by transferring the knowledge information in the source domain data to the target domain data prediction task and designing human body structure constraints and 2D pose projection constraints.Cross-domain experimental results show that the proposed method has lower prediction error than the network without knowledge transfer.(3)Design and develop a single-person 3D human pose estimation system.Based on the proposed multi-granularity feature-aligned 2D to 3D pose lifting network and unsupervised domain-adaptive 2D-3D pose lifting network,a single-person 3D human pose estimation system is designed using the Java development language.The 2D/3D human pose prediction function in images/videos are realized,and the stability and realtime performance of the system are verified by tests.