An online tracking method via improved cost-sensitive adaboost

Visual object tracking is an important problem in computer vision and has many applications including traffic monitoring, augmented reality and human computer interface. Although it has been investigated in the past decades, designing a robust tracker to cope with different objects under various situations is still a great challenging task. Focusing on the single-target tracking problem, this paper proposes an object tracking system architecture based on the simplified Haar feature and the improved cost-sensitive Adaboost with online learning strategy, and has implemented three different object tracking algorithms under this architecture. In this thesis, three main tasks have been done for the architecture to achieve the accuracy and speed required by object tracking. First, the Haar feature is simplified, for the original one's computational cost is still a burden for real-time tracking. Second, the architecture uses the cost-sensitive Adaboost instead of the original Adaboost as the object detector, because the quantity of positive samples and negative samples may be unbalanced, and the classifier will perform only as a detector without training after a certain number of the beginning frames, which will compromise the issues of real time and accuracy. Third, a new parameter is added in the sample weight updating formula of the cost-sensitive Adaboost to give more weights on the misclassified samples. In the tracking experiments, the proposed algorithm has shown strong anti-interference ability and better performance in some test video sequences, compared with other tracking algorithms. Meanwhile, the improvement of the cost-sensitive Adaboost can help the algorithm achieve better tracking results in some sequences, compared with the original one. Through experiments it can be found that the proposed tracking system architecture performs well in terms of accuracy and speed.