Human Action Recognition Based On 3D Skeletal Features

In the field of computer vision,research on human action recognition of videos is significant as well as challenging.The development of depth imaging technology,especially the advent of depth cameras such as Kinect,provides the new ideas and solutions for human action recognition.Researchers can quickly and accurately obtain the 3D positions of skeletal joints from depth images,restoring the 3D human skeleton model.Utilizing the human skeleton model,we propose two methods for action recognition on 3D human skeletal features.Respectively based on the Special Orthogonal Group(SO(3))and the Special Euclidean Group(SE(3)),this paper innovatively extracts different features of pairwise body parts in the spatial and temporal dimension.Then this paper implements corresponding feature optimization and feature fusion.Our paper uses different classification methods to successfully complete the accurate action recognition on three datasets.The detailed work of our paper is as follows:(1)This paper presents an action recognition method based on the SE(3)using feature fusion in the early stage.This method uses the matrix of SE(3)to represent the relative position between any pair of body parts,obtaining Relative Geometric Velocity(RGV).After that,this method uses feature fusion in the early stage to form the final Spatial-Temporal Moving Skeleton Descriptor(STMSD).After feature optimization,this method inputs frame descriptors into Support Vector Machine(SVM)for classification.(2)This paper presents an action recognition method based on the SO(3)using feature fusion in the later stage.This method uses the matrix of SO(3)to represent the relative rotation between any pair of body parts,obtaining Relative Rotational Velocity(RRV).After feature optimization,this method combines RRV with two other features which are Relative Joint Positions(RJP)and Joint Angles(JA)in the later stage.Aimed at three features,this method trains three SVM models.The final prediction result is obtained from the weighted fusion of three individual results.(3)This paper has made some contributions in the design of feature optimization process.The feature optimization pipeline includes three main steps which are feature processing,the temporal modeling and dimension reduction.Feature processing includes interpolation and normalization.These two aspects make samples have the same number of frames and make data standardized.The temporal modeling uses Dynamic Time Warping(DTW)or Fourier Temporal Pyramid(FTP).DTW adjusts the action sequences and reduces data noises.FTP removes high frequency coefficients to capture the action temporal structure.At last,dimension reduction uses Principal Component Analysis(PCA)to balance the accuracy and efficiency.