The Research Of Human Behavior Recognition Technology Based On The Combination Of Global And Local Features

For research of pattern recognition in the field of computer vision in recent years,more and more scholars, internet companies are willing to pay high cost into the study ofintelligent recognition technology of human behavior. Human behavior in the videointelligent identification has showed an extremely high academic value and commercialvalue. Human behavior recognition at this stage almost remain in the plain context ofsimple body movements, such as the identify of walking, running, jumping, bend,clapping, etc. If too many interference factors in the video scene, or have higher requestof real-time and accuracy of recognition, most of present algorithms are not onlysimplify,but also slowly operation and lowly recognition.The step of human behavior recognition can roughly divided into feature extraction,training features and feature classification and recognition. In this paper, after reference alarge number of international papers and advanced algorithms,we are in order to solve theissues related to the above character description. In the feature extraction step, this paperproposes a new feature descriptor PE-Cuboid which is used to characterize the humanbehavior combining local features and global features. Since there is not enough effectivefeature points to describe an action while using only single features, PE-Cuboid featurescan highly improve the recognition rate. Use pixel change ratio map(PCRM) and edgeorientation histograms(EOH) to get global movement informations, which should befurther distinguished by local cuboid descriptors. In machine learning section, the papercompares a variety of machine learning methods, it uses traditional codebook andbag-of-words, dictionary learning and sparse coding training behavior features descriptionrespectively, finally training and classifying the formed descriptions can be done by theSupport Vector Machines(SVM).It uses scoring mechanism when testing new video, thehighest score is the recognition result. This paper use matlab to simulate the algorithm. And we have tested it on somewellknown datasets, such as KTH and Weizmann, and our own datasets. The paper showhow to set parameters related to the algorithm and do a thorough analysis to comparewith the current some other advanced methods. The results show a good performance ofour algorithm.