3D Human Pose Estimation Based On Monocular Image

Three-dimensional human pose estimation based on monocular images is one of the basic tasks of computer vision.It can be applied to various fields such as virtual reality,human-computer interaction,and autonomous driving.And it has become one of the hot research topics in recent years.Since the transition from two-dimensional images to three-dimensional poses is a highly non-linear problem,and the image content contains many interference factors such as occlusion,lighting,clothing and so on,the estimation of three-dimensional human pose is very challenging.This thesis analyzes the problems for the 3D pose estimation in single-person and multi-person environments respectively,and proposes solutions.Specifically,the content and contributions of this thesis are as follows:(1)For 3D human pose estimation in single-person environment,due to the complexity of human structure,it is difficult to capture the internal relationship between all joints by using a single convolution neural network.Therefore,this thesis proposes a novel segmented 3D human pose estimation method.The human skeleton model is divided into three parts according to the correlation between the joints,and three independent convolutional neural networks are used to predict each part.In this way,the joints with strong correlation can share their own features,while the joints with weak correlation do not interfere with each other.On the premise of independent networks predicting different parts,different loss functions are designed for each network according to the structural constraints of limbs.The dynamic loss terms of symmetrical bones are added to the networks which predict upper and lower limb.These loss terms allow the network to maintain the characteristics that the length of the human symmetrical bones is basically equal,thereby improving the accuracy of the human pose estimation.The performance of the algorithm is verified on a single-person 3D pose dataset.The results show that the segmented 3D human pose estimation is significantly better than other advanced methods,and can achieve more accurate pose estimation.(2)In a multi-person environment,since the autoencoder using two-dimensional convolution can only extract the features between the same type joints of different people when performing volumetric heatmap compression,this thesis proposes a new Full Feature Extraction Autoencoder(FFEA).The autoencoder utilizes the abilities of 3D convolution to extract multi-dimensional information at the same time.When encoding the volumetric heatmap of joints,the FFEA can not only extract the features of the same type joints for all people,but also extract the features between all the joints of each person.Through extracting the multi-dimensional features,the FFEA can accurately restore the original high-resolution volumetric heatmap even at a high compression ratio,thus laying a good foundation for more accurate multi-person 3D pose estimation.(3)Aiming at different human scales in a multi-person environment,the previous methods have poor effect on the small-scale human body,and it is prone to multi-detection or miss-detection.This thesis proposes a Scale-Aware Network for Compressed Data(SANCD).The network is mainly improved on the high-resolution network in the 2D human pose estimation,and combined with the FFEA to form a complete multi-person 3D human pose estimation method—ILoCO(Improved Learning on Compression Output).The experiment results on the multi-person 3D pose dataset show that ILoCO can reconstruct the multi-person 3D pose more accurately,and the effect on processing small-scale human bodies is significantly better than other methods.