Background Modeling And Moving Object Detection Under Complex Scene

Moving objects detection of image sequences is one of the active domain ofcomputer vision and pattern recognition, which has found wide applications in robotnavigation, intelligent video surveillance system, video image analysis and so on. Thispaper focuses on the study of statistical background modeling, the foundation on whichto base the detection of moving objects.Statistical background modeling algorithms fall into two parts: parametric andnonparametric ones. Using the assumption that the background model are known,parametric algorithm turns out to be the parameters estimation. Being a typicalrepresentative of parametric algorithm, the traditional mixture of Gaussians model(MOG) is mainly dealt with. However, MOG is slowly convergent. In contrast withMOG, an improved Gaussian mixture model MOG1is proposed. MOG1introduces theforeground model and novel update mechanism, which make the objects that come torest abruptly merge into background faster. Anew algorithm for shadow elimination inthe normalized color space (I,g,r) is also proposed. Experiments show that shadowelimination in (I,g,r) is equivalent or superior to that in HSV with significantlyimproved results in case of light mutation.Moreover, a nonparametric background modeling with emphasis on the principlesof kernel density estimation(KDE) is discussed, followed by a qualitative analysis ofthe effect on the estimation of the sample size and bandwidth. From the perspective ofMISE and AMISE, the paper gives a quantitative analysis of the role of bandwidth, ageneral formula for its selection and two specific methods, namely quick and simplebandwidth selector and cross-validation.Finally, the normal kernel density estimation is described in detail. In this part, thepaper gives a normal scale bandwidth selector using quartile function. However, theselector has nothing to do with the number of samples. To overcome the disadvantage,the paper derives a quick and simple bandwidth selector with sample size as aparameter under normal density. For an arbitrary density function f, the paper gives amore universal bandwidth selector with the aid of the best approximation by twomethods for integrated squared density derivatives(R(f")), that is, convolution ofsecond derivative and kernel density estimation. For the latter, the paper puts forward anew algorithm which employs some useful conclusions under the normal density as the initial conditions, and gradually approaches R(f") with the consequence of optimalbandwidth obtainable. An online version for Ψr(g) is developed out of consideration ofcomputational complexity. Experiments show that the algorithm performs better andachieves real-time results in terms of object detection and shadow elimination.