Research On The Human Pose Estimation Method Based On The Deep Learning

As one of the basic tasks of computer vision,the human pose estimation technology aims at estimating positions of joints of human body in images and output their corresponding coordinates.In the past few years,this technology has been widely used in a series of fields,e.g.,motion capture,human-computer interaction and intelligent security.Nevertheless,in practical application,due to the influence of factors such as environmental change,human movement,clothing and so on,the images collected would suffer from the problem of missing some important features,specifically,associated information related to joints would be insufficient,which make it difficult for the algorithm to estimate positions of some joints.Moreover,the increasing of feature differences between different joints also makes it more difficult to synchronously output the precise positions of all joints.Therefore,in order to obtain effective human pose estimation results,this paper conducts systematic research from two perspectives:estimation of hard joints and feature fusion.The main work and contributions of this study are as follows:1.This study proposes estimation method of how to improve the performance of estimating hard joints based on a so called residual down-sampling module and an attention mechanism.On the basis of the Stack Hourglass network,this method mainly includes two parts:improving the estimation accuracy of hard joints and enhancing feature.Specifically,on the one hand,this paper replaces the pooling layer in the original Stack Hourglass network with the proposed down-sampling module,thus enhancing the context information in the low-resolution feature map.This method not only solves the problem of joint position information loss in the sampling process of the original network,but also effectively improves the estimation accuracy of difficult joints.On the other hand,this paper combines the attention mechanism with the residual module outside the Hourglass sub-network to avoid the interference of the added context information on the learning of effective features by the network.Basically,this method enhances the effective information in the feature map,and further improves the pose estimation accuracy of the network.According to the experimental results on the MPII and LSP data sets,the proposed method can effectively improve the estimation accuracy of hard joints and the overall estimation accuracy for a large margin.2.This study proposes a human pose estimation method based on multi-stage learning and multi-level feature reuse.On the basis of the HRNet,this method mainly includes three parts:multi-stage learning,multi-level feature and multi-stage loss function.Specifically,first of all,this paper uses the multi-stage Learning method to realize the transmission of human pose heatmaps in each stage of the network,and then effectively solves the problem of associated joint information loss problem during the process of feature fusion.Secondly,by using the multi-level feature reuse method,this study realizes the cross-stage fusion of different features in the network,and then solves the problem of shallow feature information loss during the process of feature fusion of the original network.Finally,through the multi-stage loss function,this study realizes the learning of human joint features from a coarse-to-fine way,thus avoiding the interference of noise information during estimation.In short,through the experimental verification on COCO 2017 dataset,we find that the proposed method can significantly improve the overall accuracy of pose estimation by solve the information loss problems from the above proposed feature fusion ways.