Research On Real-Time Pose Estimation And Person Identification From Multiple Views Method

Human pose estimation is a rapidly developing subject field in recent years.It is used to detect and reconstruct human joint points in pictures or videos,and has a very wide application prospect.This paper studies and improves the existing human pose estimation methods,and proposes a more accurate multi-person multi-view matching method,which can estimate the three-dimensional pose of people.This paper focuses on the algorithms of human recognition and human pose estimation,and proposes to add human recognition to the human pose estimation system.Firstly,in human pose estimation methods,this paper proposes to use depth camera and RGB camera to estimate human pose at the same time,so that the results obtained by the two devices are constrained and optimized each other.Take full advantage of the higher accuracy of color camera in non-extreme cases and the better effect of depth camera in dealing with serious occlusion.Secondly,this paper proposes to use the deep learning method to solve the problem that the accuracy of cross view matching in the previous method is not ideal.When dealing with the cross-view matching problem,the key point of this paper is to cluster the people detected from all perspectives through an algorithm based on deep learning.Each cluster encodes the corresponding relationship between the two-dimensional pose and joints of the same person in different perspectives,and the three-dimensional pose can be inferred very efficiently.Unlike most previous methods,which use Euclidean distance,the matching of the same person from different perspectives in this paper matches the same person from different perspectives through the features extracted by the feature extraction network.As for the confusion of person-ID,this paper uses the person re-identification method based on deep learning to solve it.Through the redesign of the basic network,and the person ID dataset is used for pre-training,the model has the ability to extract the features of person ID.This step reuses the basic network in cross-view matching,which greatly reduces the unnecessary calculation time of the algorithm and the memory space occupied by the model.Finally,a human pose estimation system is built,including hardware module,software module and algorithm module.The effects of various deep learning networks in this method are compared,and compared with the current mainstream methods in a variety of scenarios such as strong occlusion on public datasets and self-built datasets.The accuracy of the proposed method in self-built datasets is 92.8%,which is higher than 81.5% and 90.7% of the main stream method Mvpose and 4D Association.The method proposed in this paper solves the problem that the previous method can only identify a person temporarily,can not recognize a person who has appeared once before.