Computational Model Of The Double Channel Theory In The Visual Cortex For Action Recognition

Human action recognition is the research hotspot in the field of computer visionwith widely applications such as surveillance and human-machine interaction. It is theissue that researchers focus on that how the visual system deals with the human actionperception information and how to simulate human vision system to realize themotion recognition accuracy and quickly. With the deep research in the field of brainsciences, people know more about the human vision system. It is extremelysignificance to understand human visual information processing process deeply, andestablish recognition model which is more accurate and effective to simulate thehuman visual system. Through studying the existing bionic action recognition system,we employ the hierarchical architectures of visual system based on the existedbiologically inspired system of action recognition. The study has achieved someresults in the following aspects:Firstly, this paper presents a method for human action recognition whichcombines the form and action features. According to the basic physiologicalcharacteristics of human visual system, we proposed a calculation model to simulatethe action recognition visual cortex based on the two pathway theory. This model isbased on the original HMAX model.We add the simulation of the ventral pathway invision system, and change from analog dorsal pathway simply to simulate the ventralpathway and the dorsal pathway. At the same time，we respectively extract the formfeatures and motion features. The form pathway will extract local shape informationfrom each frame, and the motion pathway will extract the motion information. Finally,feature vectors are fused and sent to the support vector machine (SVM) forclassification. Then, we eventually realize the recognition for human action.Secondly, we put forward an effective method to obtain the feature prototypesbased on human visual attention mechanism. As the feature prototypes are importantto the recognition，we improve the quality of the feature prototypes by limited theselected areas. The specific methodology to obtain effective feature prototypes isthrough analyzing the response of complex cells to get the extract feature prototypes’locations. The system will determine the selected area depending on the recognitionpurpose, and get the useful feature prototypes. At the same time, we will obtain thesignificant region by using the visual attention mechanism and analysis the salientregion part to improve the system recognition rate.Finally, this paper presents a method for action recognition with fewer frames. For action recognition, the previous works are based on the video or most frames.After adding the shape information, we can get good recognition efficiency with lessframes and shorten the time required.This method proposed in this paper has been tested in the KTH standard database.Results show that this method improves the action recognition efficiency.