Research On Complex Human Pose Estimation Based On Deep Learning

Human pose estimation is the technology of locating keypoints(joint points)of people in images or videos and connecting adjacent keypoints.Pose diversity,illumination changes and environmental occlusion are the main factors that affect the accuracy of human pose estimation.There is an urgent application need to improve the quality of human body pose estimation in practice.This article focuses on the research of complex human pose estimation based on deep learning.After fully studying of existing human pose estimation methods based on deep learning,a new complex human pose estimation network KACNet with keypoint association constraints is proposed.The network fully considers the relationship between adjacent keypoints of the human body,and uses the relationship to assist in predicting the keypoints.A cascaded hourglass complex human body pose estimation network SH-KACNet with keypoint association constraints is proposed.The network uses the keypoint association relationship to constrain the cascading hourglass channel to achieve ke point prediction.The experimental results show that the KACNet and SH-KACNet methods can not only accurately locate and connect the keypoints of complex poses,but also have good robustness.The work of this article is mainly reflected in the following four aspects:(1)Learned traditional human pose estimation methods based on models and overall features.Several representative and latest human pose estimation methods based on deep learning have been researched deeply,and they have been implemented numerically.Point out the challenges faced by complex human pose estimation and its important value in practical applications.(2)Propose a new complex human pose estimation network KACNet with keypoint association constraints,which consists of two channels and a fusion module.The cascaded convolution channel of the KACNet network learns to obtain the keypoint position and confidence information under the constraint of the distance loss function;the keypoint association constraint channel learns to obtain the keypoint association relationship that meets the physiological characteristics of the human body under the constraint of the association loss function;the fusion module fuses keypoint position,confidence and keypoint association under the constraint of weighted loss function to obtain complex human body pose.The experimental results of a large number of public datasets and Internet data show that,compared with advanced human pose estimation methods,KACNet network can effectively suppress background interference,improve the positioning,classification and connection accuracy of keypoints of complex human pose,and has good robustness.(3)The cascaded hourglass complex human body pose estimation network SH-KACNet with keypoint association constraints is proposed.Compared with KACNet,SH-KACNet obtains more keypoint category labels in its keypoint association constraint channel,which helps to improve the prediction accuracy of the cascaded hourglass channel.(4)Enhanced LSP dataset,and constructed an enhanced dataset EN-LSP suitable for training complex human pose networks.Improve the special evaluation index OKS of MS COCO dataset,and design the evaluation method OKS_m that does not rely on the dataset.