Object Tracking with Collaborative Extreme Learning Machines

We propose a novel collaborative discriminative model based on extreme learning machine (ELM) for object tracking in this paper. In order to represent the object more precisely, we first propose a new collaborative discriminative representation model, which includes both a global discriminative sub-model and a local discriminative sub-model. Different from traditional local representation models, in particular, our local sub-model integrates several classifiers which have structural relations to improve the expression. The global discriminative model represents the appearance comprehensively while the local discriminative sub-model can effectively address occlusions and assist the update. Second, to have better combination of these sub-models, we propose a novel collaboration strategy based on the Kullback-Leibler (KL) distance. The novel strategy can determine the weights of the sub-models adaptively by measuring their KL distances reciprocally. Third, we introduce ELM into tracking and adopt it to build both the global and the local discriminative sub-models simultaneously. Since ELM has a good generalization performance and is robust to the imbalance of the training samples, it is suitable to be used for tracking. Experimental results demonstrate that our method can achieve comparable performance to many state-of-the-art tracking approaches.