Learning Algorithm for Real-Time Vehicle Tracking

This article presents a learning algorithm for real-time object tracking in video sequences which uses an improvement of a feature selection method known from object detection. But in contrast to trackers based on object detection methods, our approach explicitly selects the features which are best suited to track an object, which are different from the best features for object detection. The used features are constructed from pairs of image patches and related to Haar features. Besides the automatic selection of features according to their discriminative (tracking) power, the advantage of this approach is that the resulting tracker is very fast, allowing it to run in addition to a detector to robustify the object position estimation and to compensate for dropouts of the detector. A comparison of the proposed tracking algorithm with other tracking methods is presented which shows the accuracy of the proposed algorithm