Haar-Like Feature Based On-Line Boosting Tracking Algorithm with Directional Texture Entropy

Large numbers of features are needed for tracking in the conventional Haar-like feature based on-line boosting methods (HBT). The vast amount guarantees the existence of useful features. But it leads to expensive computing and memory requirements. On the other hand, the coexisted useless features may bias the tracking towards random influences. To address those two problems, in this paper, we propose a new method named HBTDTE (Haar-like feature based on-line Boosting Tracking with Directional Texture Entropy). In this novel algorithm, the texture direction and richness information are extracted by the entropy of gray level co-occurrence matrix (GLCM). Then Haar-like features are only calculated along directions with richest texture. And the employed feature number is also governed by the texture richness. In this way, not only the number of useless features is largely reduced, but also the total size of the feature pool. Experiments reveal the higher robustness and performance of this new HBTDTE method. We show the superiority of our method over several current state-of-the-art tracking methods in several experiments on publicly available data.