Research On 2-D Human Pose Estimation Based On Deep Learning

As a key step in dealing with human activities,human pose estimation mainly involves the computer locating the joint points of characters from images or videos.Human pose estimation can be divided into traditional graph model method and deep learning method.The graph model method relies on prior knowledge to cope with complicated attitude transformation.The method based on deep learning does not rely on the prior knowledge of the model and can achieve better results.Human pose estimation can be divided into two types: two-dimensional and three-dimensional according to dimensions.Because threedimensional estimation can be obtained by reasoning with two-dimensional pose estimation,two-dimensional pose estimation has important research value.Most of the two-dimensional human pose estimation algorithms based on deep learning have the following problems: only the convolution features of the last layer are used,and the connection between convolution layers is ignored,which is easy to cause information loss;Only the local features of human joint points are considered and the global features are ignored,resulting in low recognition accuracy of joint points.It is easy to misjudge joint points and limit their accuracy by regression of joint point coordinates or classification prediction of joint point thermal diagram.For multi-person analysis,the traditional method matches joint points from all candidate joint points,which easily leads to wrong connection.In this paper,the above problems of human pose estimation are studied.The main research contents are as follows:(1)A single human pose estimation algorithm based on multi-level dense block hourglass network is proposed.The algorithm is mainly composed of the following parts: Firstly,feature extraction is carried out by using dense blocks,because each layer of dense blocks is connected with each other,which greatly reduces the loss of feature information of the whole network;Secondly,the hourglass unit is used to realize multi-resolution feature recognition of human pose,i.e.first from high resolution to low resolution and then from low resolution to high resolution.This network structure of feature sampling and fusion on multi-resolution can capture and integrate human body joint information of different scales.Finally,a relay monitoring mechanism is used for each hourglass unit to alleviate gradient disappearance.(2)An algorithm based on simultaneous regression of joint offset and thermal diagram is proposed to realize the research of multi-person attitude estimation.The algorithm is mainly composed of the following two stages: the first stage encoder layer,for joint point detection,classifies and predicts the thermal diagram of joint points and the regression coordinate 2-D offset vector at the same time to realize more accurate joint point location;for joint point association,it uses component association fields and has the ability to store fine-grained information on low-resolution activation maps;In the decoder layer of the second stage,Hopcroft-Carp algorithm is used to analyze the human posture.Hopcroft-Carp algorithm transforms a K-map matching problem into a bipartite matching,which can greatly improve the accuracy and reduce the time complexity.In order to verify the performance of the two models proposed in this paper,they are trained and tested on FLIC,MPII and MS COCO data sets respectively.The experimental results show that the accuracy of hourglass network based on multi-level dense blocks is higher than that of several other single-person pose estimation algorithms,especially for small joints,such as elbow and wrist,which are difficult to predict.The average accuracy rate and recall rate of multi-person pose estimation algorithm based on deep learning are higher than those of other mainstream algorithms,which can effectively reduce false detection and missed detection of joint points.