Local features classification for adaptive tracking
Chapter in Scopus
- Additional Document Info
- View All
In this paper we propose using invariant local features and a global appearance validation for assisting a robust object tracker initialized by a single example. Although local features have been used before for several object detection and tracking applications, these approaches often model an object as a collection of key-points and descriptors, which involves constructing a set of correspondences between object and image key-points via descriptor matching or key-point classification. However, these algorithms cannot properly adapt to long video sequences due to their limited capacity for incremental update. We differentiate from these approaches in that we obtain key-point-to-object correspondences instead of key-point-to-key-point correspondences converting the problem into an easier binary classification problem, which allows us to use a state-of-the-art algorithm to incrementally update our classifier. Our approach is embedded into the Tracking-Learning-Detection (TLD) framework by performing a set of changes in the detection stage. We show how measuring the density of positive local features given by a binary classifier trained on-line is a good signal of the object's presence, and in combination with a global appearance validation it yields a strong object detector for assisting a tracking algorithm. In order to validate our approach we compare the tracking results against the original TLD approach on a set of 10 videos. © 2013 Springer-Verlag.