Abstract
We propose a novel algorithm to identify occlusions for visual tracking through learning with observation likelihoods. In our technique, a target is divided into regular grid cells and the decision for occlusion is made for each cell using a classifier. Each cell in the target is associated with many small patches, and the patch likelihoods observed during tracking construct a feature vector, which is used for training and testing. Since the occlusion is learned with patch likelihoods instead of patches themselves, the classifier is generally applicable to any videos or objects for occlusion reasoning. Our occlusion reasoning algorithm has decent performance in accuracy, which is sufficient to improve tracker performance significantly. The proposed algorithm is combined with an $L_1$ minimization tracking algorithm, which is also used for training to simulate the patterns of likelihoods with and without occlusions. The tracking outputs as well as occlusion reasoning results are illustrated in many challenging videos, and both qualitative and quantitative evaluation are performed to present the advantages of our algorithm.
