The Study Of Deep Learning Based Human Pose Estimation

In recent years,the role of human pose estimation in areas like security and human machine interaction is increasing.Visual information occupies a large proportion in human activities,and human body movements are the most intuitive and easy to obtain in visual information.Besides,human pose estimation can be embedded into video information perfectly and it is necessary for the understanding of human action in video.Therefore,human pose capturing and effective estimation becomes important,especially nowadays,various of visual equipment and mobile applications promotes the understanding of human pose estimation.Due to the problem of blur,different scales,different human pose and inconsistent resolutions for the pictures from visual equipment,it is hard to deal with those problems simultaneously.In addition,some other factors from video aggravate the understanding of human pose.In order to estimate human pose more accurately,this paper propose a human pose estimation method based on the deep convolutional neural network,and introduce human pose estimation from a single picture into video to realize the human pose tracking in video by combining the relevant technologies such as people re-identification and human detection.The main contribution of this work are:(1)A cascade pyramid network for human pose estimation in single frame image are proposed in this paper.Based on the deep convolutional neural network,this algorithm divides the whole network into two steps: global network and refine network.On the basis of the residual network,global network directly uses the structure of feature pyramid to fuse the different image feature which has different scales so as to achieve the rough estimation of human pose and to avoid making multi-scale pictures as input of global network.Based on global network,the refine network has some different treatments for the output of the feature pyramid,and concatenate them at last to refine the above rough estiation with L2 loss.At the same time,it uses online hard example mining technology to learn those “hard” keypoints.Essentially,the whole network training is end to end,because it uses intermediate supervision with the combination of global network and refine network,and it is beneficial to learn the entire network by making two networkscomplement each other.The experimental results show that the proposed method is very effective and it achieves the state-of-the-art results.(2)We propose a human pose tracking algorithm using video data.The algorithm uses a local alignment based people re-identification technology as appearance model to extract features of the human body,and exploits a single object tracking algorithm to solve the problem of multi-object tracking innovatively.The single object tracking algorithm can not only generate the tracklet,but can also compensate the missed video frames and improve the smoothness of the tracklet.Unlike the multi-object tracking algorithms,they just make use of a series of rules designed to concatenate small tracklets into full trajectories.In specific implementation,we not only the use feature extracted by the people re-identification,but also use the IOU between adjacent frames for similarity measure.Both of the two measurements take advantage of the spatial-temporal information of video repeatedly to make the algorithm more robust.The final experimental results show that the proposed method is effective and competitive.In this thesis,sufficient experiments are demonstrated by MSCOCO and Posetrack datasets which are designed for human pose estimation,and the experimental results show that the proposed method is reasonable and effective.