Online Detection of Concept Drift in Visual Tracking

In the field of data mining, detecting concept drift in a data stream is an important research area with many applications. However the effective methods for concept drift detection are seldom used in visual tracking in which drifting problems appear frequently. In this paper, we present a novel framework combining concept drift detection with an online semi-supervised boosting method to build a robust visual tracker. The main idea is converting updated templates to a data stream by similarity learning and detecting concept drift. The proposed tracker is both robust against drifting and adaptive to appearance changes. Numerous experiments on various challenging videos demonstrate that our technique achieves high accuracy in real-world scenarios.