Study On Action Recognition Using WEB Data

The rapid development of Internet technology makes everyone can become a Web data contributor.Web data is not only closer to the real human action scene,but also contains a lot of data which is increasing every moment.How to use these Web data to achieve human action recognition effectively and automatic is now a hot issues of intelligent video surveillance video retrieval and intelligent human-computer interaction.The traditional action recognition method can only train a more robust classification model when the training data is sufficient.However,the video in the real world contains all kinds of actions.A new action is not able to get a good classification model when training samples are few.Given the maturity of commercial visual search engines,Web data may be the next important data source to scale up visual recognition.As multimedia search engines mature,it's possible to get low-cost tagged data from the web.The introduction of a large amount of Web data can effectively solve this problem.Moreover,a large amount of Web data can be used as auxiliary data to enhance the performance of the target domain.Web action images usually portray characteristic scenes for an action,which highlight discriminative portions of a video's temporal progression.So this is a strong evidence that Web images can enhance action recognition.Clearly,there are complementary benefits between the temporal information available in videos and the discriminative scenes portrayed in images.We propose an algorithm which can enhance action recognition by using a large number of Web images.We get a lot of Web images of related action on the network,And then we put the Dense trajectory of the video together with the characteristics of the Web images into the SVM.We reduce the data distribution mismatch between Web images domains and video domains by using cross-domain dictionary learning algorithm.Because the Web images can be easily obtained on the Web,the proposed algorithm can enhance action recognition with at almost zero cost.Experimental results show that the proposed algorithm can improve the accuracy of human action recognition effectively on KTH and YouTube sports action datasets.