Action Recognition System For Still Images

Human action recognition is one of the most important and popular topics in the field of computer vision.It has a widely usage,such as the anomaly detection in public,the analysis of consumer behavior,video indexing and monitoring,as well as image search.To different input,the methods of action recognition can be divide into two different kinds,the video based action recognition and the image based action recognition.Although,it is almost twenty years from the appearance of first action recognition method,the first image based action recognition algorithm was designed in 2006.Recently,caused by the variance of pose and gesture of human and the complication of the background and scene,most image based action recognition algorithms are complex and performance are not so good to be used in practice.There are two kinds algorithms in existing image based action recognition methods,the global image representation and the part representation of human body based action recognition.Specifically,the global representation methods not only try to describe the human pose for action recognition but also combine the information of scene of the action.However,this kind of methods can't distinguish some actions that only very small difference,cause the representation they give are not include the detail information of human pose.While,in the other methods,they give the feature which extracted from the key part of human body,so that the representation can give the details of human pose.But,this kind of methods may ignore some information of scene.Considering the complication of image based human action recognition,this paper proposed a presentation method that can both depict the human pose and describe the interaction between human and object.Here,the human is the person who is doing the action,and the object is the thing that have a strong interaction with the person and can be used to describe the action.In most cases,the interaction is the spatial relationship between the human and the object.Based on this idea,we first design a part based representation methods,which can describe the human pose by combining the information of human body parts.Also,we describe the interaction between the human and the object by depicting the overlap area,the distance,the relative location of them,Lastly,we combine the representation of human pose and the interaction,and using a classifier to classify the action of the human.In particular,we evaluate our method on two public action image datasets,the results show that the proposed algorithm has powerful capabilities in depicting the human action and classifying the human actions.