The Reseach On Visual-based Human Abnormal Behavior Recognition Algorithm

As a hot research issue in the computer vision, human abnormal behavior recognitionresearch has been widely used in various fields. It has the vital practical significance forimproving and enhancing the quality of people’s lives. However, due to the complexity ofhuman action itself, and the influence of the environment, many difficult problems in theresearch of visual-based human abnormal behavior recognition need to be solved. Therefore,the focus of this thesis is mainly on the research of feature extraction and detection algorithm,finally good results of abnormal detection are achieved.The difficulty in the description of human motion is to achieve reliable and stabileinformation. Firstly, a method of feature testing is adopted to get the basic feature, and moreadaptive and reliable spatio-temporal interest points feature is selected. The3-dimensionalscale-invariant feature transform (3D SIFT) is employed to describe these interest points.Then a novel method is proposed to describe human behavior. The positional distributioninformation of the interest points is extracted and combined with the base feature. Owing tothe high dimension of original feature, twice principal component analysis are done on thefeature from single-frame and multi-frame to reduce dimension. And then the feature isfurther quantified and selected to remove the redundant information, so as to focus on thedescriptor containing critical motion data. Finally, AdaBoost iterative algorithm combinedwith the traditional support vector machine is used to detect abnormal behavior. The keytraining samples are selected by using AdaBoost, and the final good strong classifier isconstructed by weak classifiers of each layer. Experiments are carried out on the publicdatabase of KTH, and the results show that by using the proposed feature and detectionalgorithm, the correct detection rate of abnormal behavior has been significantly improved.In this thesis, through the further research on action feature and detection algorithm, thebehavior description is more accurate, simple and stable. As a whole, the proposed algorithmobtains satisfied performance for human abnormal behavior recognition.