Object Tracking is one of the main research directions within the field of computer vision, which has significant applications. There are a variety of methods to solve tracking problem, a very important one of which is classification-based approach. That object tracking based on classification trains a discriminative model and updates online. In this thesis, we focus on four algorithms in the framework of boosting. In the exist way, weak classifier is picked that minimize loss function each round, and therefore reduce train error. However, with a large number of experiments, we found that randomly choosing a certain amount of weak classifiers improves tracking result.Our works can be concluded as follows: (1)Review four tracking algorithm based on boosting framework. They are online MIL, and three other algoritms based on boosting framework, online AdaBoost, online Gentle AdaBoost and online SavageBoost. Also, an improvement has been made on the problems in MIL and SavageBoost. (2)Through large number of experiment on different video clips of different tracking algorithm, we found that randomly choosing a certain proportion of weak classifiers can lead to more stable tracking. The experimental results show that in application, the best proportion of randomly chosen classifiers is subject to specific tracking algorithm and video sequence.
修改评论