Study Of Multispectral Video Streams-based Human Action Recognition Algorithm

This paper systematically studies the multispectral video streams-based human actionrecognition algorithm and the realization of the key technologies which has importantacademic value and practical significance for human behavior recognition technology and theapplication and development of intelligent video surveillance technology. This paperdiscusses some key problems of action recognition and further studies the relevant keytechnologies of action recognition and then put forward a novel method of action recognitionwhich obtains a good recognition performance.This paper analyzes the research progress of action recognition technology both at homeand abroad, studies the feature extraction and modeling method in action recognition,establishes the human behavior recognition detection model based on the ordinary light andthe active infrared respectively, proposes the behavior recognition model based on decisionfusion in order to realize the human behavior recognition under the multispectral condition.The main works of this paper includes:â‘ The basic theory of video streams-based human behavior recognition issystematically studied. Most of the behavior recognition researches based on video are focuson the visible area currently. In order to realize the behavior recognition in the environment ofno light or weak light, scholars and researches have explored to analyze the behavior underthe infrared light. With the progress of science and technology, more multispectral calculationis used from military to science research. Based on the researches above, the multispectralhuman action recognition model is proposed which integrates the common behaviorrecognition characteristics under ordinary light and infrared light and can accurately identifythe human behavior under the ordinary light or infrared light in order to realize therecognition of the all-weather behavior.â‘¡The human action recognition algorithm model under the ordinary light is putforward. The paper uses the composite feature of3DHOG-HOOF which is a blend of threedimensional gradient direction histogram features and optical flow gradient directionhistogram feature. Through the fusion of the two different characteristic features, the videopixels in the three-dimensional distribution of local area and the feature descriptor of pixelschange have been combined in order to improve the accuracy of behavior description of thehuman body.â‘¢The human behavior recognition algorithm under the active infrared light is put forward. The paper proposes the composite feature of3DHOG-DFC which is a blend of threedimensional gradient direction histogram features and the feature of distance between thefeature spot and the centroid. The paper also describes the external morphology andcharacteristics of human actions in order to represent the state of the human behavior of videomore accurately. This method has a good recognition effect under the condition of activeinfrared.â‘£The behavior recognition decision fusion model based on D-S evidence theory underthe multispectral condition is designed. By using the rule of Dempster, the recognition resultsin the behavior recognition framework both under the ordinary light and under the activeinfrared are made into a new belief function as the final decision in order to achieve thehuman behavior discriminates under the multispectral condition in a more believable level.This behavior recognition decision fusion model has high flexibility which can effectivelyreflect the different sides of and different types of information. It also has the advantages offault-tolerant and strong anti-jamming capability.