Human Action And Event Recognition Based On Privileged Information

In recent years,human action and event recognition has become an important area of computer vision.It is widely applied to the fields of human-computer interaction,video surveillance,health care and video retrieval.It has brought convenience to human production and life,and has broad application prospects.Robust recognition method is the key to its large-scale application.In this paper,we propose two robust human action and event recognition framework,namely:human action and event recognition framework based on privileged information,human action and event recognition by domain adaptation sparse coding suing privilege information.The classical human action and event recognition method extracts consistent visual features of the training and test samples to learn a classifier.In this paper,we propose a framework of human action and event recognition based on privileged information,extract robust privilege information only for training samples to learn SVM+ classifier.As for human action recognition,this paper uses the human skeleton depth feature as privileged information and verifies its effectiveness in two public datasets,namely UTKinect-Action3D and Florence 3D Actions.As for human event recognition,this paper describes the characteristics of Web textual descriptions as privileged information,and verifies the robustness of the algorithm on the Flickr dataset.In the classical human action and event recognition,the training and test samples have the same feature space and common data distribution.However,it is difficult to satisfy the above conditions,and collecting labeled training samples is often time consuming and expensive.In this paper,we propose a method of human action and event recognition by domain adaptive sparse coding suing privilege information.This method allows training and test samples from different fields.To minimize the distribution differences between the source domain and the target domain samples,we combine domain adaptation with graph sparse coding,and we also take advantage of the privilege information in the source to learn a better classifier.The algorithm achieves good results in the experience where Flickr dataset as source domain,CCV and Kodak dataset as target domain respectively.