Research On Object Tracking Algorithm In Intelligent Video Surveillance

With the improving of the hardware equipment performance, intelligent videosurveillance system is wildly used in the security monitoring field. Moving targetstracking is the significant module of which. In the case of targets fast movement,appearance and illumination changing and background disorder achieve the movingtargets tracking is the key points and difficulties which the researchers concerned.In many targets tracking algorithms, our research concentrates on the more usualmean drift algorithm and particle filter algorithm in the intelligent video surveillancesystem area. Mean drift algorithm is an algorithm based on the nonparametric densityfunction estimation, which convergences to the function peak along the biggestdensity function gradient change direction by the iteration method, such that get theposition of targets tracking. Calculated amount of mean drift algorithm is less andeasy to achieve it, which is commonly used in real-time targets tracking. But to thebackground disorder and targets small situation, the mean drift algorithm is morecommonly lost the targets. The particle filter algorithm achieve the recursive Bayesianestimation of the target state through the monte carlo simulation method, such thatextract the random state particles from the posterior probability of target state, andcorrect weight of each particles by observing the similarity between the target andtemplate. Eventually in conjunction with each particle to represent target statedistribution, and thus determine the target state. Particle filter algorithm represent thetarget state by statistic average of multiple particles, which still can get a more robusttracking result as the noise influence is more significant. But there exists drawbacksof more calculated amount and sample lack in the particle filter algorithm. This paperchapter2analyses theories and describes the algorithms in detail of above twoalgorithms.The main research contents and innovation is：(1) Due to the background disorder and targets small situation in monitoringscene, the existing algorithms can hardly realize robust tracking, we put forward atargets tracking algorithm based on background perception, Which introduces thetarget partial background to target description and represents the targets as a point setwith weight. Then calculating the observation probability of tracking frame, combinedwith the prior information to get the target posterior probability value of the searcharea, then implanting to the mean drift algorithm frame to achieve target tracking.Experiment proved that our algorithm is robust to the complex circumstancementioned above.(2) Toward to appearance changing as the target moving, the existing target tracking algorithm is just to create a single reference model for the target to describe,the target reference model can hardly adapt to the target appearance change that willlead to target drift even lost tracking. We proposed a particle filter tracking algorithmbased on multiple reference models, which builds a multi-color histogramcombination for multiple prior appearance of the target, after convex combinationoptimization and calculating the optimal combination coefficient, we get the optimalreference model and realize moving target tracking under the particle filter frame. Theexperiment proved that our algorithm can achieve a good result in tracking theappearance changing targets.