Visual Object Tracking with Pyramid, Random Subspace Features

In the classical visual object tracking, the observation model and the inference model are two essential parts for efficient tracking. Although the strong discriminative features have been used to model observation model in tracking, most of these features have high dimension, which handicaps the tracking speed. In this paper, we introduce a pyramid random subspace to build the observation model, which can efficiently utilize the high-dimensional feature. Concretely, the pyramid feature is used as an instance of the high dimensional feature. The random projection (RP) is used to reduce the dimension of feature at each pyramid level. The matching adopts the weighed strategy over the all pyramid level. Finally, the observation model is incorporated into particle filtering to inference the state of the visual object. The extensive experiments are carried out to show the accuracy and efficiency of the proposed method, under the condition of variation of illumination and pose variation.