Tracking via Discriminative Online Learning of Local Features / Learning Features for Tracking
We present an approach treating the problem of object tracking as a matching problem of detected keypoints between successive frames. The novelty is that discriminative classifiers for local features are learned online. This allows a simplification of the classification problem among the currently detected keypoints. Furthermore, different to existing approaches, we are capable to start tracking of objects from scratch meaning no off-line training phase is needed. An on-line boosting technique is used for learning a distance function to distinguish currently detected keypoints. Samples are collected as new frames arrive making the classifiers more and more robust over time. This allows to distinguish learned local features from each other. A simple mechanism incorporates temporal information for selecting stable features. The approach can be used within real-time applications since on-line updating and evaluating classifiers can be done very efficiently.
