Semi-supervised visual object tracking by label propagation

Recently, object tracking is viewed as a foreground/background two-class classification problem. In this paper, we propose a non-parameter approach to model the observation model for tracking via graph, which is a semi-supervised approach. More specially, the topology structure of graph is carefully designed to reflect the properties of the sample's distribution during tracking. In predication, the confidence of sample's label is propagation via random walk with restart (RWR), which can utilize labeled or unlabeled samples easily. The primary advantage of our algorithm is that it keeps the appearance of object in graph model, which can easily model the multi-modal of object appearance. Experimental results demonstrate that, compared with two state of the art methods, the proposed tracking algorithm is more effective, especially in dynamically changing and clutter scenes.