Online learning for classification and object tracking with superpixel

Visual tracking is an important task in computer vision. Treating object tracking as a binary classification problem has been already discussed in recent years. State of the art classification based trackers perform better robustness than many of the other existing trackers. In this paper, we consider object tracking as a binary classification problem. A Random Forest classifier is trained on-line based on superpixels to distinguish between the object and the background. The classifier is then used to label superpixels in the next frame as either belonging to the object or the background. A confidence map is formed from the classification scores. The tracking task is then formulated by finding the peak of the map, where is the position of the object. In order to locate the position, an improved mean shift is proposed to work on the map. We show a realization of this method and demonstrate it on several video sequences. Experimental results show that our method is capable to handle heavy occlusion and recover from drifts.