Research On Deep Learning Based 3D Human Pose Estimation

At present,3D human pose estimation is one of the most popular research areas.It is aimed at estimating human joint positions from images or videos,and assembling to a full human pose.3D human pose estimation is the base of human pose recognition,action recognition,human tracking and other tasks,and as the same time,it has great application value in rehabilitation,video monitoring,human-computer interactive and other application fields.In recent years,with the achievement of deep learning based on neural network in various application fields,deep neural networks with powerful learning capabilities have gradually become the best choice for some complex tasks like 3D human pose estimation,and lots of novel researches prove the performance of deep learning once again in 3D human pose estimation.In this thesis,we aim at some problems like occlusions and interference in 3D human pose estimation,and we propose two improved methods to the network architecture and geometric constraint.1.For the 3D human joints regression in 3D human pose estimation,the spatial image features are not taken into consideration properly.To further capture and utilize the inherent spatial image features in different resolutions,we propose to use a multi-scale recalibrated approach to improve the accuracy of estimating 3D human pose,and the recalibrated branches are integrated into the network through transfer learning.Both quantitative and qualitative results demonstrate the effectiveness of our model,and we also achieve a competitive result on Human3.6m dataset,which is the largest public 3D human pose dataset.2.When we are estimating some special human poses,the optimizing of the network improves the performance of these bad cases negligibly.So,an advanced geometry constrains model for the human pose is proposed in this thesis,and this model is introduced to a semi-supervised learning method.The model restricts the human pose from joints position,bone length,and joint angle.A validity analysis on loss function is performed after the integration of geometry constrain model into the loss level.Some quantitative and qualitative comparative evaluations are shown and analyzed.