| One of the fundamental problems in motion analysis is using similarity measures to make meaningful comparisons.This is of great importance in movement teaching,movement standard degree evaluation,and abnormal behavior detection.With the rapid development of science and technology,computer vision technology has been greatly improved.Among them,motion capture technology is used to digitalize human motion data,which lays a solid foundation for motion analysis.With the development of this technology,we can analyze and compare motion more accurately,which provides a better possibility for the realization of various tasks related to motion.In recent years,analyzing human motion data has become one of the research hotspots in computer vision.However,when analyzing and comparing the motion data,the description of the human body structure may change with the change of viewpoint,and the skeleton expression of the same posture captured from different viewpoints is quite different.On the contrary,the current 2D attitude detection methods can achieve high performance.This paper will consider the use of 2D data to analyze human motion,focusing on how to identify the similarity of human motion in multi-view.(1)This paper aims to make a meaningful comparative study on how to use only the 2D pose of the human body,as this has not been fully explored in existing work.Current 3D human pose estimation based on 2D RGB images is still imperfect,and requires more performance than 2D human pose estimation methods.To this end,this paper proposes a human pose embedding network,which is used to learn the viewinvariant feature of human pose across views.The network maps the 2D pose to the view-invariant Gaussian embedding space,and outputs the mean and variance of its Gaussian Distribution.The similarity of the corresponding distributions of different 2D poses represents the matching probability between their real poses,and solves the uncertainty of 2D poses.This method can measure pose similarity across views.(2)This paper studies lightweight human pose estimation,because the embedded model uses 2D human body key points as input,but the current 2D pose extractor greatly increases the complexity and computation of the model due to its high resolution,and cannot be used in resources such as mobile.Running on a limited device slows down the efficiency of analyzing and comparing motion poses,limiting the popularity and application of this technology.In order to solve these problems,this paper first proposes a lightweight module to replace the basic module in the highresolution network,which greatly reduces the amount of parameters and calculation of the model,and then constructs a feature fusion module based on the attention mechanism for the multi-branch structure,which further improves the model’s feature extraction ability and ensures that the pose estimation task runs efficiently without reducing performance.Finally,a large number of contrastive experiments are carried out for the proposed models.Experimental results show that for 2D posture estimation tasks,the proposed lightweight model achieves significant performance and achieves the same performance as other large networks.For the study of human motion across views,the embedding model proposed in this paper also has a good performance,which proves the effectiveness of the embedding attitude.In addition,in order to speed up the popularization of the technology of human motion comparison,a motion-matching system is implemented based on the model mentioned above. |
