A Biologically Inspired System Based On Attention Mechanism For Action Recognition

Understanding the perception of actions in humans is an important area ofresearch crossing the boundaries between several scientific disciplines from computerscience to brain science and psychology. Motion recognition is one of the mostchallenging technologies in video processing system with important applications suchas surveillance and human-machine interaction.Because of the difference between motion character and the condition to recordvideos，action recognition is not very easy in computer vision. But at the same time，our understanding of the brain mechanisms responsible for the recognition of actionshas progressed over the past decades and biologically inspired systems for actionrecognition growing rapidly.On one hand，these biologically inspired systems can giveus a more clearer picture of human visual mechanism，on the other hand，they are newmethods of action recognition in computer vision. This thesis addresses thehierarchical architectures of visual system based on the existed biologically inspiredsystem for action recognition. The study includes some aspects as follows:This paper analysis a biologically inspired system based on attention mechanismfor action recognition. To solve the shortcomings of already existed recognitionsystem which has a lot of computation and time consuming，we bring the attentionmechanism into the recognition system. To make use of HAMX model，our systemcan locate the motion object exactly.This paper also presents a model of spatio-temporal saliency visual attention.Because Itti’s model can not eliminine the influence of background noise，we makeuse of the motion energy extracted from videos to get the foreground information.Through this way to reduce the influence of background noise，and obtain a moreeffective system.Feature prototypes are important to the recognition result，the performance ofthese feature prototypes determine the motion feature’s excellence which extract fromvideos. So we present a method to get effective feature prototypes without missingaccuracy and speed. The specific methodology to obtain effective feature prototypesis through analyse the response of complex cells to get the extract feature prototypes’slocations.Finally，we test the approach on different publicly available action datasets such as Weizmann human action dataset and KTH human action dataset. The experimentalresults show that our method can improve model’s rate of identification rate and speed.In addtion，the results of cross validation between different human action datasetshows our system has strong robustness.