Research On Multi-person Pose Estimation Based On Convolutional Neural Network

In the field of Computer Vision,such as Intelligent Monitoring,Image Retrieval,Human-computer Interaction,and Augmented Reality,human activities receive much concern,the human pose estimation has become an indispensable part with urgent applications requirements and is also a very popular research direction in the field of computer vision.Human pose estimation refers to detecting the positions,calculating the directions and scale information of various parts of the human body from static images or sequence images.Due to the diversity and flexibility of the human body,small changes in any part of the human body will also generate new attitudes.In the meantime,the visibility of keypoints of the bone is affected by factors such as posture,viewing angle,occlusion and background.In recent years,Convolutional Neural Network has been effective in visual tasks.Therefore,human pose estimation based on deep learning is a very challenging subject in the field of computer vision,and it also has important research significance.This thesis is based on the deep convolutional neural network structure to study the human pose in static images.First of all,unlike top-down or bottom-up research methods,in order to improve the accuracy of target detection,this thesis uses a top-down research method as a whole to performs human target detection and initial keypoint estimation in static image based on Mask R-CNN,while training the keypoints of the human body in combination with the keypoint mask.At the same time of detection,the task of initial estimation of keypoints is completed.Secondly,in order to improve the accuracy of bone keypoints estimation,the research method of PersonLab(used for bottom-up human pose estimation)is borrowed,which strengthens the constraints before bone keypoints.Finally,keypoint constraints are fused to the keypoint estimation,and output with target detection of the final experimental results on human pose estimation.In the course of the experiment,this thesis firstly performed human target detection on the CrowdHuman dataset,and proved the feasibility of the network in target detection.However,it does not involve key point information.And in this case,we performed human keypoint estimation on MSCOCO data set which has a larger number of targets,and also compared the experimental results of different CNN structure research methods in that data set.Experimental results show that this method can further improve the accuracy of human pose estimation in static images.