Research On Human Action Recogniton Method In Video

Video-based human action recognition enables computers to understand human behaviors from visual data.As an important part of computer vision,there have been active researches on this topic because of its wide applications.Given that human actions are complex,diverse and ambiguous,so it still remains to be a challenging task for automated understanding.In this thesis,we work on these main themes towards solving this task as follows:motion foreground get,visual interest point detection,feature modeling,including global feature and local feature,moreover,choice of action classifier.In this thesis,we propose a back-and-forth frame difference method to abstract motion foreground.On this basis,we eliminate noise according to connected domain area and it's distance to rare foreground image center.Compare to traditional two or three frame difference method,proposed technique get more complete and accurate motion contour,moreover its high efficiency outweighs the so called Gauss mixture model in application.This thesis also put forward a human motion region based global feature detection and represent method,by extracting minimum enclosing rectangle and circle of motion region,we represent it as a five tuple,including rectangular degree,length and width ratio,circularity,inclination and eccentricity ratio.In this paper,we propose a local interest point detection method based on moving body parts.Three adjacent frames difference output two pieces of motion information of current image,after which deploy FAST(Features From Accelerated Segment)feature point detection,then intersection of the two points set is taken as final output with non-maximum suppression.With the low time-consuming FAST algorithm applied,this method should be an efficient motion interest point detector.In this thesis,we also explore the popular BOW(Bag of Words)model for human action recognition.K-means and AP(Affinity Propagation)cluster method are used to generate visual words,which is also called codebook.we find K-means cluster based visual word generate method get better performance.We explore the reasonable visual vocabulary size of BOW model,and also find that the traditional visual vocabulary generation method does not take into account the weight of the vocabulary.To overcome this weak point,we propose an improved BOW model based human action feature vector generation method effectively use the association between words.The more times a visual word comes in the final human action representation,the less contribution it do to classification,so we weight the visual words according to its cluster center density inversion,effectively improves the recognition accuracy.Furthermore,different feature detectors,descriptors and classifiers are compared.The classifier used is SVM(Support Vector Machine),KNN and Decision Tree.Performance is tested on different datasets,the simple KTH dataset,Weizmann dataset and more complicated UCF50.Cuboids based descriptor get better performance than patch based descriptor,SVM classifier get best accuracy but KNN is more efficient.