Learning With Collaborative Local Disturbance For 3D Skeleton-based Human Action Recognition

Machine learning has been widely applied in a variety of domains,which usu-ally depends on a large amount of data.However,in practice,the existence of high-dimensional data brings difficulty to many learning tasks.To overcome such chal-lenges,many subspace-based methods have been proposed.By representing the high-dimensional data as low-dimensional subspaces and dealing with these subspaces,the problems could be addressed.Among these subspace-based methods,learning on Grassmann manifolds has attracted much attention due to their high consistency in han-dling subspaces.However,it neglects that subspace representations could be disturbed by noise in the original data.Meanwhile,traditional subspace representations might ignore the specific type of data and eliminate too much information,which probably leads to a reduction of accuracy in discriminative learning.This paper revisits Grassmann-based discriminative learning and makes a further discussion with regard to the above issues.Based on sequence data,this paper pro-poses a new approach,collaborative local learning with Disturbance Grassmann ker-nels,which simultaneously considers potential subspace disturbances and temporal in-formation to obtain more robust and accurate classifiers.Moreover,it yields a new family of collaborative local disturbance kernels and leverages kernel methods for dis-criminative learning.The application of skeleton-based action recognition is used as a running example throughout this paper.And empirical experiments have been con-ducted on several synthetic or real-world data,which verifies that the proposed method could achieve good performance in many situations and on different sub-tasks.The main contributions of this paper include:1.To better analyze the task of action recognition under noise,this paper proposes a newly collected dataset based on real-world human actions.Different from others,the dataset contains several noise actions specified for noisy action recog-nition;2.To deal with subspace disturbances,a systematic introduction to learning with dis-turbed subspaces is provided in this paper.It extends learning with marginalized corrupted features to nonlinear space and noise is added for data augmentation,which improves the robustness of subspace learning;3.This paper proposes a subspace representation based on product Grassmann man-ifold which intrinsically embeds the local temporal information of sequences.It compensates for the information loss during dimension reduction and as a result enhances the accuracy of subspace learning;4.In this paper,collaborative local learning with disturbance kernels is proposed.It effectively employs disturbances and redundant information,which improves a tradeoff between the accuracy and the robustness of action recognition.5.Experimental studies are conducted on multiple datasets,which proves the supe-riority of the approach.