Study On Tracking Algorithm Based On Data Mining Theories

After comprehensive conclusion and detailed analysis on theadvantage and disadvantage of the tracking algorithm in existence andfull understanding of Concept Drift and Class Imbalance in the fieldof Data Mining, this paper propose an novel object tracking algorithmwithin the framework of Concept Drift, which make the best use ofthe probability knowledge of change-point detection to update andadapt the tracking algorithm. The algorithm described in this paper isembedded in the framework of Particle Filter, which achieves a goodperformance even when concept drift occurs. We demonstrate theperformance of the algorithm in through two typical scenariosexperiments: the scenario with light mutation and the scenario withinwhich the motion pattern of the object changes. The main work of thispaper is as follows:Having propose a effective concept which works in object trackingfield. We know that the study of Concept Drift is confined in DataMining, and we extend Concept Drift to Computer Vision andexplain how to obtain a effective concept.Using Bayesian Online method to detect change-points. TheBayesian theory is a hot spot in Data Mining and we find that itfits the probability in tracking algorithm which lay the foundationfor us to unite them in the same probability framework.Having proposed a novel tracking algorithm based on ConceptDrift. In this process, the knowledge of concept drift is embeddedin the posterior likelihood probability function based on ParticleFilter. So, once the Bayesian Online method find change-pointsoccur, it will be reflected in the tracking model. This is the reasonwhy the proposed algorithm can adapt the changing scenario fast.In addition, we also discussed the problem of Class Imbalance intracking algorithm which is often exists with Concept Drift at thesame time. Generally speaking, Class Imbalance is likely to occur inthose tracking algorithm based on classifiers, the disproportion ofpositive and negative samples introduce a big influence to the trained classifier. The main idea of the algorithm proposed in the last sectionis: we introduce the CI module to the training process of the classifierwhich helps in promoting the classification accuracy and trackingperformance.