Research On Complex Human Pose Estimation Via Keypoints Association Constraint

Human pose estimation(HPE)is a technology for estimating skeletal keypoints of the target person from images/videos.It is urgently needed in the fields of human computer interaction,medical treatment,fatigue monitoring and entertainment.Limited by the difference in shooting angles,the instantaneous movement of the target person and the accuracy of the equipment,HPE still faces many difficulties.Therefore,HPE based on deep learning has important theoretical and practical application value.This thesis focus on improving the accuracy of two-dimensional(2D)complex HPE based on deep learning methods.Based on the research of advanced HPE methods based on deep learning,we propose a double keypoints association constraint network DKAC-Net,which effectively improves the estimation accuracy of complex human poses.We further added the direct and indirect prior association between keypoints into the DKAC-Net to obtain the EDKAC-Net.EDKAC-Net improves the ability of keypoint’s association constraint,which further improves the accuracy of complex HPE.The main work of this thesis is reflected in the following three aspects:(1)We have studied the classic and deep learning-based HPE methods.Additionally,we have numerically implemented several representative methods based on deep learning.We have described the HPE frameworks based on deep learning,the current difficulty of complex HPE and the application value of the research on HPE.(2)We propose a complex HPE network DKAC-Net with double keypoints association constraint.DKAC-Net is composed of the feature extraction module,the keypoints prediction module and the fusion module.It effectively predicts complex human poses under the constraint of double keypoints association with physiological characteristics.Extensive experiments results show that the DKAC-Net method can not only effectively improve the classification and positioning accuracy of keypoints of complex human poses,but also has good robustness.(3)We propose an enhanced keypoints association constraint network EDKAC-Net that introduces the direct and indirect prior association among keypoints into the DKAC-Net.EDKAC-Net applies Hough transform to dataset labels,and uses the direct and indirect prior association between the keypoints to enhance the prediction ability for keypoints of the DKPAC-Net,which more effectively improves the accuracy of complex HPE.