Visual tracking via an ensemble of random classifiers
Yichun Shi, Hesheng Wang · 2016
One largest problem for tracking-by-detection methods is the incomplete and noisy training set. Occlusion, illumination and many other problems could lead to this problem. Models that are not adaptive enough would fail to track the target when drastic appearance change takes place. Adaptive ones, although keep tracking at first, could lose the target because of learning too many incorrect samples. In this paper, we present an ensemble model of random classifiers updated on different dataset. When the appearance of the target changes drastically and some sub-models are confused, the others could help correct the tracking result. A latent variable is added for choosing sub-models and it naturally leads to predicting the new samples with a weighted sum of sub-models. To calculate the weight, we add a generative model to each random classifier. Experiments show that our method could track the target robustly and accurately.