Research Of Target Tracking Algorithm Based On Online Learning Model

Target tracking is a hot and difficult research in the field of computer vision. It hasbeen widely used in the military and civilian fields.Target tracking technology based ononline learning at the same time of tracking update training classifier, can better adapt to thechange of the target, and not only for one type of target tracking.Because of the obviousadvantages of online learning tracking technology, it quickly became the research emphasisin the field of tracking.But some popular online learning tracking algorithms are still easyto drift, be easy to degenerate and have badly real-time performance.So this article based onsome existing algorithms discusses the target tracking technology based on online learning.This paper has been mainly completed the following several aspects work.(1) Introduce the basic principle and implementation of the existing tracking algorithmbased on online learning. The compressive tracking (CT) algorithm and the TLD(Tracking-learning-detecting) algorithm are mainly discussed and analyzed the advantagesand disadvantages of the algorithms.(2) CT algorithm can not choose effective compression features after these features aregenerated.So this paper proposes a solution based on effectiveness of features.Byconstructing feature pool, these features will be sorted in the pool and the most effectivefeatures are selected as a classifier to use.The experiments on public data sets proves thesorting method based on normal distribution distance has stronger adaptability and stability.(3) In this paper, multiscale CT algorithm based on the optimized genetic algorithm(Multiscale-Genetic Algorithm-Compressive Tracking, MGACT) is improved in theoriginal CT algorithm.Firstly, the MGACT uses the genetic algorithm to select the optimalfeatures and improve the effectiveness of the features and real-time performance of the CTalgorithm.Secondly, duing to the searching boxes size of CT algorithm cannot be changed.This paper thorugh improving the expression of compressive features solves the problemthat the value of features changes with the size of searching box changes.Kalman filter andconservative classifier are introduced to solve the problem of the object shelter and littleprevious information classifier reserved of object.Thus through testing some public datasets, the MGACT achieves the better tracking performance.(4) As the real-time performance of MGACT algorithm is not perfect, and the effect ofMGACT algorithm is not obviously increased when the illumination is changing quickly in the environment.The paper puts forward MGACT algorithm based on the framework ofTLD which improves the speed and accuracy of the algorithm.Through experiments aretested on some public data sets, algorithm efficiency is obviously improved in the lightchange quickly, etc. A pedestrian and several pedestrians can be robustly tracked in amonitoring environment.