Research On Human Action Recognition

In recent years,the study of human action recognition has become a hot and difficult topic in the field of computer vision.The purpose of human action recognition is to automatically identify human action category from a sequence of images or videos,which is significant for achieving intelligent surveillance,human-computer interaction,virtual reality and it is an important issue to promote the practical application of computer vision technology.This paper firstly describes human action recognition method by using the current popular methods,based on spatial-temporal interest point description methods and also introduced method of vision feature representation which is based on low-level features' statistics and attribute learning model based on semantic information.Then the two main research contents are discussed in detail: 1)human action recognition method based on multi-feature fusion.To address this problem,we use multi-instance learning method for learning and integration through various features of the human body movements to take full advantage of the characteristics of the various features,so that we can get richer information of the human action to help identify human action.2)human action recognition based on hierarchical feature.During the current study of human action methods to extract essential features,however,most of the features are artificially designed to extract information from data,which may miss the characteristic information.Through using the hierarchical feature learning from the smallest spacetime blocks,we can combine the neighbor information of the blocks to make full use of spatial and temporal correlation information around the blocks.Finally,we introduce two human action datasets and evaluate above methods on the two datasets.The experiment results were analyzed in detail.