Research On Human Pose Estimation Method Based On Stepped And Multi-Residual Module Convolutional Neural Network

The human pose estimation task is to determine the precise pixel position and feature information of various parts of the human body in an input image,such as arm,torso,leg and head,or each joint points such as elbow,wrist,knee and shoulder.It is a basic and challenging task in computer vision tasks,and is also an important foundation for advanced visual tasks,such as action recognition,human-computer interaction,costume analysis,re-identification recognition,and motion recognition.And widely used in autonomous driving,virtual reality,video surveillance and animation.Effective human pose estimation must be able to deal with rare human poses and images with large limb changes,without the influence of clothing diversity,limb flexibility,lighting conditions and complex backgrounds,while also addressing serious challenging problems such as limb deformation and severe obstruction of body parts or joints.Most of the traditional human pose estimation is based on the pictorial strictures models and combined with some feature information,such as texture features,color histogram features,shape context features and gradient histogram features to construct the human body's appearance model and part relationship model.Although these methods can locate the position of human joints roughly,they are vulnerable to the interference of complex background,and the prediction of some human parts with large deformation or severe occlusion are still inaccurate.In recent years,deep learning has made breakthroughs in the field of human pose estimation.In particular,the powerful learning ability of deep convolutional neural networks(DCNN)has attracted more and more researchers.Its wide use provides a strong non-linear transformation capability for human pose estimation methods.In order to estimate human pose more effectively,this thesis deeply studies the deep convolutional neural network,and builds a stepped deep convolutional neural network for human pose estimation,which improves the accuracy of pose estimation;A multi-residual module stacked hourglass network is designed for the traditional stacked hourglass network,which further improves the accuracy and robustness of human pose estimation.The main research work and results of the thesis are as follows:1?A human pose estimation method based on stepped deep convolutional neural network is proposed.In order to impro-ve the accuracy and robustness of human pose estimation,the network first uses the step module as building block to explore the interdependence between human body parts in human pose estimation.Secondly,in order to capture more and more context information,the network increases the effective receptive field of the network output layer by using the step module of the large convolution kernel.Finally,the network enhances the intermediate supervision by learning the last loss function of each step module,thereby supplementing the gradient of back propagation,effectively solving the problem of gradient disappearance during training.The experimental results show that the proposed human pose estimation method based on stepped deep convolutional neural network improves the accuracy of human pose estimation to a certain extent,and also reduces the influence of occlusion problem on pose estimation.2?A human pose estimation method based on multi-residual module stacked hourglass network is proposed.In order to further improve the accuracy of the human pose estimation,this paper aims at the influence of the dimension change of human body parts caused by the change of camera view on the final human pose estimation accuracy in two-dimensional human pose estimation.Firstly,an improved residual module is designed based on the traditional residual module.Then,in order to learn the correlation between human body parts in the large receptive field,a large receptive field residual module is designed.After that,in order to better solve the inaccuracy of pose estimation caused by the variation of the scale of various parts of the human body,a multi-scale residual module is designed in this paper.Secondly,these three residual modules are used as the building blocks of the hourglass sub-network.When the resolution is high,the large receptive field residual module and the multi-scale residual module are used to capture the information of a larger range and scales.When the resolution is low,only the improved residual module is used.Finally,four multi-residual module hourglass sub-networks are used to form the final multi-residual module stacked hourglass network.The experimental results show that the proposed human pose estimation method based on multi-residual module stacked hourglass network further improves the accuracy of human pose estimation.