Research On CNN Based Keypoints Estimation

Human pose estimation is the task of predicting the coordinates of keypoints of hu-man body from images.It is the basis and premise of some more advanced visual tasks,and has a wide range of application scenarios,such as human-computer interaction,intel-ligent photo editing,monitoring and other fields.In recent years,human pose estimation has become a hot research direction in the field of computer vision.Nowadays,researchers have put forward some good effect of two dimensional hu-man body attitude estimation algorithm,but the existing algorithm has not output point visibility/keep out information,even if the use of the information is only used to help im-prove the precision of prediction of key,and the keypoints of visibility/block information for downstream tasks such as pedestrian recognition,gesture recognition is very impor-tant.In this thesis,the visibility knowable two-dimensional attitude estimation algorithm is studied,which not only predicts the coordinates of keypoints,but also predicts their vis-ibility.In order to evaluate the effectiveness of visibility prediction,this thesis proposes an evaluation index for the accuracy of visibility prediction.Furthermore,a top-down and a bottom-up visibility knowable human pose estimation network is proposed.The main contributions are as follows:(1)An index to measure the accuracy of visibility prediction of keypoints is proposed.(2)A top-down visibility-aware human pose prediction network is proposed.Differ-ent network architecture designs are tried and compared.Using the CrowdPose dataset as our benchmark dataset,the visibility accuracy reached 91.3%,and the additional pre-diction tasks helped improve the precision of the keypoints,the MAP,by 0.8 percentage points.(3)A bottom-up visibility knowable human posture prediction network is proposed.In the middle layer the feature predicts the visibility tag and adds supervision,and then uses the visibility heat map to help with the prediction of the human instance tag.Visibility information can be effectively predicted with almost no loss of positioning accuracy of keypoints.(4)An optimization network for human body pose estimation is proposed,which can effectively improve the result of keypoints location.