Research On Parameters Reflecting The Tracking Precision And Optimization Of Tracking Mode

Video target tracking refers to analyzing, detecting, extracting, identifying and tracking forthe interested object in the video sequences, has been used in the human-computer interaction,navigation of robots, and medical diagnose, etc, and is the basic of realizing intelligent videosurveillance. In real monitoring scenes, it is difficult to track the target precisely for somefrequently occurring instances, such as the occlusion, the variation of the state of the movingtargets, illumination variation and dynamic variation of background, etc. While the trackingalgorithms presented are specified solutions for given situations, finding out the algorithm whichcan use in any environments is still difficult.This research puts forward such an opinion, which is integrating all the tracking algorithmspresented up, ranking them from easy to difficult, and monitoring each of them by differentparameters. Then switch over to the next algorithm while the algorithm used cannot track thetarget precisely. The paper here takes the auto-switch between the Kalman Filter and Particle Filteras an example.Contraposing the random and indeterminate elements, such as the occlusion, the variation ofthe state of the moving targets, illumination variation and dynamic variation of background, etc,which can affect the tracking precision, are conducted by noise. Here define Gaussian Degree tothe random noise, and divide them into four kinds, who are super-Gaussian、Gaussian、sub-Gaussian and non-Gaussian. Then find out the range that Kalman Filter can work and defineadaptive-parameter to increase the tracking precision of Kalman Filter when the noise obeys tosub-Gaussian.Proposing a parameter that can evaluate the performance of tracking algorithms, the wayused considers the frame prior to the local one, which is different from the usual pattern that isused in video processing. It is not only pledging the continuous of message, but also evaluating theperformance of algorithms. While put it into use in Kalman Filter, it can be used as the parametermonitoring the sequential switch between Kalman Filter and Particle Filter, once exceeding thethreshold, switch Kalman to Particle Filter. Take the improved Gaussian Degree as the parameter monitoring the opposite sequentialswitch between Kalman Filter and Particle Filter. The improved Gaussian Degree is obtained bythe difference between the sample mean and the sample on the basis of Gaussian Degree. Theswitch threshold is got by the sum of sample on the basis of improved Gaussian Degree. Thisfinally leads the realization of switching Particle Filter to Kalman Filter.The mutual switching mode between Kalman Filter and Particle Filter makes it possible thatintegrating all the tracking algorithms presented and dynamicly choosing the best one in certaincircumstances.