Scale and occlusion invariant tracking-by-detection

In this paper we present a solution for tracking-by-detection that is able to handle both scale variations and occlusions of the tracked object. We build upon the framework structured output SVM tracking and improve it in order to deal with both variations of target scale and occlusions. We first propose to modify the original solution to include the scale variations both in the patch sampling stage and in the structured output state. Then in order to deal with occlusions we introduce an incremental classifier to discriminate the target from the context. This classifier combines a learning phase with a unlearning one that help to avoid drift in the model of the tracked object.