Research On Filting Algorithms In Visual Tracking

Filtering technology is a key technology in target tracking.It can estimate the targetstate recursively,which is a important guarantee to get a precision output in a trackingsystem. The typical filter algorithms such as kalman filter is adaptive to linearsystem,and the improved algorithms such as extended kalman filter or unscentedkalman filter is used in nonlinear system. But their filtering performances will descendor even diverge when non-gaussian distribution occurs. Particle filter algorithm can beimplemented by a bayesian recursion process though a monte carlo simulation method.It has great advantages in non-linear or non-gaussian fields.In this thesis,first,some basic filtering algorithms such as kalman filter,extendedkalman filter and unscented kalman filter were introduced,and the advantages anddisadvantages of the various algorithms were discussed through simulations.Then,thisthesis proposed an improved resampling algorithm to solve the problem of particledegeneration. The systematic resampling was used to improve the multinomialresampling algorithm. The results of simulation showed that the average performance ofthe improved resampling algorithm is superior to the other Resampling algorithms'.Atthe same time some improved PF algorithms are studied,the main focus was analgorithm called unscented kalman particle filtering algorithm based on minimal skewsampling.This thesis also analysed the advantages and disadvantages of every improvedalgorithm through simulations.At last,SIR-particle filter and unscented kalman particlefilter algorithm based on minimal skew sampling were integrated to solve the problemof large amount in calculation with the latter algorithm in visual target tracking.Thesimulations showed that the improved algorithm increases the operating efficiency by25% without precision drop.Besides,in this algorithm it is easy to achieve a differentbalance between filtering accuracy and run time by a simple parameters adjustment.Soit can adapt to the different requirements for the filter algorithm in different systems.