Single online visual object tracking with enhanced tracking and detection learning

Single online visual object tracking has been an active research topic for its wide application on various tasks. In this paper, a new framework and related approaches are proposed to solve this problem consisting of enhanced tracking and detection learning. In the enhanced tracking part, an appearance model based on correlation filter with deep CNN features and a dynamic model using improved pyramid optical flow method are employed. Two models cooperate together to depict object appearance and capture target trajectory, which also contribute to provide training samples for detection learning. In the detection learning part, a cascade classifier and P-N learning scheme are employed to reinitialize tracking when model drift occurs. Data experiments on several challenging benchmarks show that the presented method is comparable to the state-of-the-art.