Several Methods On Background Subtraction Techniques For Moving Objects Detection

In the field of computer vision, detecting moving objects is one of the most basic tasks,whichwill greatly impact the subsequent work such as classification, tracking and high level under-standing of events and so on. There are two classic methods on the detection of moving objects,optical ?ow method and background subtraction method. In video based surveillance with fixedcamera, method based on optical ?ow can still detect the moving objects even in the situationthat the camera shakes due to wind or something else. However, as it is necessary to calculatethe optical ?ow which is computing consuming and has low anti-noise ability, it is seldom usedfor Real-time processing. So, at present, the most-studied and most-extensively-used methodis still based on background subtraction.The basic idea of background subtraction is that we compare every incoming frame witha reference frame. If the property of the pixel value from current frame presents a certaindiscrepancy to that from the reference frame at the same position, we will determine the pixelpoint to be foreground point in the current frame.The main work of this article includes the following two aspects:Firstly, we introduced several traditional background subtraction techniques, for Example,method based on temporal di?erence, method based on temporal axis filtering and methodbased on Single Gaussian Model and so on. And at the same time we gave their advantagesand disadvantages.Secondly, we highlighted the detecting methods used under complex scenes .As it is knownthat the Mixture of Gauss Model theory is not systematically analyzed, and the variant al-gorithms are not comprehensively compared with each other, on the basis of reading a lot ofreferences, we analyzed mixture of Gauss through systematically theoretical derivation.As toKernel Density Estimation,we gave comprehensive introduction from theoretical analysis to itsapplication in background modeling in video surveillance, summarized its advantages and disad-vantages objectively, and made a profound analysis on its calculation. This article also discussedthe background subtraction technique based on clustering, detailedly introduced a backgroundmodeling algorithm based on codebook construction using vector quantification and clusteringtechniques,and compared it with Mixture of Gauss and Kernel Estimation from several aspects. Background subtraction technique based on Mixture of Gauss Model assumes that thetrue background pixel value at a fixed point should obey multiple Gaussian distributions, So itestablishes the probability density function for the observed pixel value at this point aswhereμk,σk2 separately represents the mean value and the covariance of the kth Gaussiandistribution ,andωi the corresponding weight .For the current observed value Xt of this point,the algorithm determines that from whichdistribution it most likely comes by calculating which Gaussian distribution has the largestposterior probabilitywhere P(k|Xt,Θ)means the probability by which Xt comes from the kth distribution of themixture model. At the same time,based on certain rules,the algorithm estimates which dis-tributions can be used to represent the background, as a result,all Gaussian are divided intotwo parts,B and F.Only those in B are thought to be able to represent the background. Soby judging whetherXt belongs to B or not,we determine Xt come from background or fromforeground. In addition,the algorithm updates the parameters of the mixture model using anincremental EM algorithm,so that its adapts to dynamic scene.Background subtraction technique based on Kernel Density Estimation treated the ob-served value X1,X2,...,XN from the same point of the latest N frames as the samples.Usingthese samples it estimates the probability density function of the current background distribu-tionwhere K(x) is the kernel function ,andσi is the smoothness parameter and specifies the"width"of the kernel around each sample point Xi. For any current observed value Xt,we calculate thedensity estimation f?(Xt),and compare it with a certain threshold T. If it is smaller than thisthreshold, we determine that it comes from foreground.The codebook method based on clustering has the background e?ectively compressed andencodes it into codewords.All these codewords form a codebook ,which is an e?ective represen-tation of the true background.For an observed value Xt(Xt may be intensity or RGB vector),thealgorithm considers its distortion to the codeword both from color dist and intensity dist.And if this distortion is within certain range,we determine Xt comes from the background that thiscodeword represents. If Xt comes from none of the backgrounds that the codewords in the bookrepresent, then it is determined to come from the foreground ,otherwise from background. Inaddition,the algorithm also adopt aλ-filtering technique as well as a layered modeling strategyto update and maintain the codebook.