Multiple kernel boosting based tracking using pooling features

We present a novel Multiple Kernel Boosting (MKB) based tracking method using pooling features. Pooling features are a type of features generated by pooling methods, which are more distinctive and significant than low level features. In the pooling step, we not only use average pooling and spatial pyramid max pooling, but also propose the pyramid random pooling, which can pick the activation within each pooling region. First, the features obtained from the poolings are used to train weak single kernel SVM classifiers, which then are combined through MKB into a strong classifier. Numerous experiments on challenging datasets show that our tracking framework achieves promising.