Research On Techniques Of Moving Object Detection Under Complex Environment

Moving object detection is one of the most important tasks located in the bottom of computer vision systems,which has widely application in many domains such as Intelligent Video Surveillance,HCI(Human Computer Interaction)and multimedia application.However,in complex application scenario,because of the dynamic background,a gradual and sudden illumination changes,shadow and camouflage,and other factors,moving object detection technology is facing great challenges.In this dissertation,we focus on moving object detection methods under the complex scene.The contribution and innovation points of the thesis can be summarized as follows:(1)Aiming at the problem of existing background subtraction algorithm is sensitive to illumination changes and shadow for complex scenarios,a novel sample consensus background subtraction algorithm based on LSBP feature is proposed.The novel based on singular value decomposition LSBP features can effectively restrain the effects of light and shadow and noise,at the same time for more flat area also has distinction.Algorithm combines color features and LSBP texture features,use sample consensus based two levels authentication,and dynamically adjust the thresholdin accordance with the complexity of the background.The experimental results demonstrate that in complex scenarios,the method can deal with illumination changes and shadow better than single color features detection methods.(2)Aiming at the problem of fixed learning rate in the traditional gaussian mixture model in complex dynamic scenarios result in background update not in time,the dissertation presents a new mixed gaussian background model based on classification of complex scenes.Background pixels divided into heat background(dynamic background)and cold background(static background)two types,and LBP and LSBP double texture feature classification method is used to distinguish true and false foreground,finally,specify different learning rate for different pixel types Qualitative and quantitative experimental results demonstrate that the proposed algorithm can improve the robustness and adaptive background model effectively,processing sudden illumination changes,dynamic background,image noise etc.(3)Aiming at the problem of existing batch mode low rank approximation algorithm is lack of real-time,the dissertation proposes a moving target detection method based on online low rank approximation.This method retains advantages of the batch DECOLOR algorithm and online ORPCA algorithm,improve the problem of these two algorithms,uses the sparsity and connectedness terms of DECOLOR method and estimates the background model using sequential low-rank approximation of OR-PCA.In foreground estimation stage,using MOG model isolated real foreground from outliers,therefore solved the the effects of image noise and dynamic background.The experimental results demonstrate that the proposed method realizes the online real-time and reliable motion object detection,and can achieve the performance of the batch algorithm and response of the dynamic background and image noise effectively.