A Novel Approach of On-line Discriminative Tracking Feature Selection

We propose a framework of a multi-object tracking method based on image classification. Aiming to build an appearance model of a target, a sparse measurement matrix is used as a projection for dimensional reduction and online feature selection to improve discriminative power. Our main idea is putting appropriate weights on the features selected to form a feature pattern. Using our method, it is easy to build a multiobject tracking framework, the tracking performance of which is dependent on the power of classifiers. Since our feature selection method is motivated by experience and from the observation of experimental results, we discuss why our method works well. Experiments show good performance of our method. Keywords-Sparse Random measurement, Naive Bayes classifier, Feature Selection, Multi-object Tracking