Object tracking via combining discriminative global and generative local models

Abstract

In this paper, in order to track objects which undergo rotation and pose changes, we propose a novel algorithm that combines discriminative global and generative local model. Initially, we exploit the wavelet approximation coefficients and completed local binary pattern (CLBP) to represent the object global features. With the obtained global appearance descriptor, we use online discriminative metric learning to differentiate the target object from background. To avoid the drift problem results from global discriminative model, a novel generative spatial geometric local model is introduced. Based on SURF features, the generative local model quantizes the geometric structure information in scale and angle. Then, we combine these global and local models so that they can be benefit each other. Compared with several other tracking algorithms, the experimental results demonstrate that the proposed algorithm is able to track the target object reliably, especially for object pose change and rotation.