Moving Object Detection And Behavior Recognition In Intelligent Visual Surveillance

In recent years, Intelligent Visual Surveillance technology has become one of agreat hot research area in computer vision, also is the focus in the field of computerapplication. Intelligent Visual Surveillance system used visual computing technology,video image processing and pattern recognition, artificial intelligence technology torealize automatic surveillance by image processing, analysis and understanding and therelated research results had been showed a great effect on social benefit and economicbenefit in the intelligent traffic management, public security management,human-computer interaction technology, construction of smart city construction,medical, and many other aspects.This thesis started from the intelligent visual surveillance of theoretical researchand practical application, introduced the research background of intelligent visualsurveillance, then also through the presentation of the development of intelligent visualsurveillance and progresses of intelligent visual surveillance in domestic and foreignresearch, finally the main research contents are presented as follows:1) In the area of the object realtime detection, first, the state-of-the-art algorithmsin background modeling method were analyzed, and an improvement was made onGMM model, the combination of background subtraction and temporal differencing, butmainly focused on the codebook mode. Based on the limitation of the original codebookmode which is lack of the adaptive ability and update efficiency, an improved codebookmodel using time series statistics was proposed. In the proposed algorithm, in order toimprove the probability of matching the active code word at the first time, a new rapidmatching method was used to sort the codes at the top of the code words according totheir hit times, also a new cache update strategy was applied to replace the originalupdate method which update the code words frame by frame. Experiments showed thatthis method not only can eliminate the effect of dynamic factors, but also robust tohandle the problem of multiple moving objects in illumination change dynamic scene.2) In the area of the object realtime tracking, aiming at the defects of the traditionalcolor histogram mean shift tracking algorithm that spatial information of the target wasnot contained, and the tracking inaccurately lost when the object color closes to thebackground color, this thesis proposed a novel tracking algorithm with blob feature,which determined blobs method by the size of the least enclosing rectangle and usedhorizontal projection and vertical projection of moving targets to determine the movingobject blobs. Blobs determined their weight coefficient by the Bhattacharyya coefficient. Also the Kalman filter was applied to estimate the blobs position, effectively optimizethe productivity of mean shift tracking algorithm. Through the experiment that aconclusion could be made that the proposed method could adaptive to determine thepartitioning method of the object and have well tracking performance than thetraditional mean shift algorithm under some cases such as partial occlusion anddeformation.3) In the area of the moving object behavior analysis, a iming at effective featureselection for human behavior classification problem, a key frame-based featurematching method is proposed. Based on a fact that human motions were a posture set intime sequence, and through study of the periodicity of the human motions, a summarizeis made between the least enclosing rectangle of human motions and key-posture indiverse activities. A hybrid wavelet moment is employed to establish key-posturetemplate, and similarity between each galleries and key-posture template was calculatedbased on that the hybrid wavelet moment matrix distance was token as the similaritystandard of the image, and recognized motion category. Experiment results showed thatit could realize the classification of different speed action within the same period of timeand increase the precision of hunman behavior classification.4) Based on the analysis of key-posture, a hybrid wavelet moment was applied ashuman action features representation, and the improvement of wavelet contour momentparameter selection would reduce the noise interference, also would improve theproblem of the motion feature redundant and highly in computational complexity. Onthe basis of markov property of the human body movement sequence, the humanbehavior HMM (Hidden Markov Model) was established. By comparing the similaritybetween observation sequence and key-posture template, a hybrid wavelet momentwhich has minimum distance between observation sequence and key-posture templatewas used to get the encoding as input of HMM; According to the Baum-Welch trainingof hidden markov model algorithm, the observation sequence was used to recognize andclassification the human identification by means of the Viterbi algorithm．Experimentresults showed that this method has reached the higher recognition rate at complexbehavior recognition, and had good versatility for different types of behavior.5) Using the Behavior matrix generated by HMMs study, a combined NaiveBayes classifier was proposed to solve the human behavior class ification. Each staticposture was treated as a state, and these states were linked through their probability,then human behavior could be a state set of static posture. Therefore human behavior classificiaton could be achieved by calculating the maximum state probability betweendifferent static postures set through traversal. Experiments showed that the methodshows a better recognition performance, application and check performance, which wasmore suitable for the requirement of complex environment.