Robust Tracking Integrated with Detection and Online Learning

This paper focuses on long-term target tracking which need defining bounding-box in the first frame, and then performing target tracking automatically. Observing that neither tracking nor detection can solve long-term tracking task independently, we combine tracking with detection module into a single framework to perform robust tracking. An online learning step is integrated into the framework to update the target model and the training samples to further increase system robustness. Our algorithm is evaluated quantitatively and qualitatively which can achieve saturated results.