Robust Visual Tracking Algorithm Based on Structural Multi-scale Features Adaptive Fusion in Co-training

As traditional visual tracking algorithms based on co-training framework are characterized by poor robustness in numerous real-time scenarios, a robust visual tracking algorithm based on structural multi-scale features adaptive fusion in co-training is thereby proposed. In order to achieve a comprehensive representation, we select structural multi-scale features in gray intensity space and local binary pattern (LBP) space respectively to train the corresponding Bayes classifiers. Online feature selection criterion via maximizing posterior probability entropy function is utilized to avoid constructing a oversized weak classifier pool. The collaborative strategy based on confidence score distribution entropy of partial candidate samples is utilized to measure the weight of each classifier. In addition, to relieve the updating error of co-training accumulation problem, we bring up a novel co-training criterion with cascaded HOG classifier as an assistant. Comparative experiments on extensive challenging sequences demonstrate that the proposed tracker is of better performance than other similar or state-of-the-art trackers.