Research On Global And Local Information For Video-based Action Recognition

In recent years,with the rapid development of computer science and technology,computer vision,which is a new field of study,has gradually become a hot issue.Compared with computer image,the form of video is more intuitive,and its information is more abundant,so it is widely used in the rapid development of the current multimedia.Under the era of big data,intelligent recognition not only promote the understanding of the human visual system itself,but also meet the needs of artificial intelligence applications.Automatic recognition of massive video has become an urgent need in the field of computer vision,and the action recognition is critically important to the key technology of video content understanding and analysis.It's vital to the research,wide application in the field of robot navigation,video surveillance,intelligent transportation,human-computer interaction and virtual reality.Feature extraction and representation are the key to the success of video content description and subsequent video processing.Therefore,this paper studies the method of video action recognition based on feature extraction and representation.Recently,the research of action recognition based on feature extraction and representation mainly uses features such as shape,color,motion to describe the content of video,but there are some limitations.For example these features can describe human action posture,but can not distinguish the action of similar pose in the video.At the same time,they are usually able to describe the video data with simple background and single target,but can not deal with the video data with complex background and action.In order to solve these problems,this paper proposes an action recognition method which bases on global and local information.The method integrates multiple visual features to describe the video content.The global information reflects the overall information of video which can maintain the time series of video.In the meanwhile local information reflects local details of the information.The combination of global and local information can effectively distinguish the similar postures of the video,which can also deal with the background in complex and multiple targets of video data.In this method,the global information is mainly based on motion feature of video which uses multi-scale optical flow histogram features.Local information analyzes the shape and structural relationships of the local area,and uses the space-time shape context feature and 3D gradient direction histogramfeature.For the combination of multiple visual features,there are two feature fusion mechanisms: feature level fusion and decision level.This method was applied to the action recognition task through the experiment in the open standard video databases.The action recognition effect is better when we compare with classical recognition algorithms.It's important that we verify the validity of this method.