Based On The Filter For Moving Target Tracking

Moving target tracking in video sequence is the core issue in thefield of computer vision. Filtering theory based moving target trackingalgorithms, because of their unique advantages, are widely used invideo surveillance, vision navigation of robot, military guidance andmany other fields. Although people has carried on the extensive andin-depth research in visual tracking, and many effective solution havebeen proposed, there are still many problems to be further discussed andstudied with respect to visual tracking under complex scene.On the basis of the analysis of particle filter algorithm of and kalmanfilter,the fusion algorithm and the effectiveness of adaptive adjustmentscale factor in unscented kalman filter have been discussed in thisdissertation. The main work and contributions are in the following fouraspects.(1)The theses has analyzed the existing tracking algorithms,emphasized on expounding of the importance sampling method and theparticle filter algorithm, discussed the relative merits and framework ofthe algorithms.(2)In order to solve the problem of expensive calculation in particlefilter algorithm, a fusion method is proposed which combines particlefilter algorithm with mean shift algorithm. The mean shift algorithm isapplied to preliminary target tracking, and tracking results are comparedby threshold. When the tracking results is not satisfactory, particle filteralgorithm is conducted to improve tracking performance.(3)The basis of the Unscented Kalman Filter method is expounded,some derivations,sampling strategy and basic algorithms are alsointroduced. The paper also gives a analysis of the main factorsinfluencing the UKF Filter performance.(4)In order to overcome the problem of artificial setting scale factor in the UKF filter algorithms, the paper has proposed an adaptiveadjustment method for scale factor. By using error between the predictedvalues of UKF nonlinear approximation and the corresponding realvalues, scale factor is adjusted adaptively and sampling strategies arecorrected. In this way, the practicability and intelligence of the algorithmare improved, and the accuracy is guaranteed.