Object Tracking via Dynamic Feature Selection Processes

We propose an optimized visual tracking algorithm based on the real-time selection of locally and temporally discriminative features. A novel feature selection mechanism is embedded in the Adaptive Color Names [2] (ACT) tracking system that adaptively selects the top-ranked discriminative features for tracking. The Dynamic Feature Selection Tracker (DFST) provides a significant gain in accuracy and precision allowing the use of a dynamic set of features that results in an increased system flexibility. Our ranking solution is based on the Inf-FS [6]. The Inf-FS is an unsupervised method, it ranks features according with their “redundancy” (for further details on the original method see [5]). For the sake of foreground/background separation, we propose a supervised variant that is able to score high features with respect to class “relevancy”, that is, how well each feature discriminates between foreground (target) and background. Therefore, we design the input adjacency matrix of the Inf-FS in a supervised manner by significantly reducing the time needed for building the graph. The ACT tracking system does not fit the size of the bounding box. Indeed, in the original framework, the bounding box remains of the same size during the tracking process. We propose a simple yet effective way of adapting the size of the box by using a fast online algorithm for learning dictionaries [4]. At each update, we use multiple examples around the target (at different positions and scales), we find tight bounding boxes enclosing the target by selecting the one that minimizes the reconstruction error. Thus, we also improved the ACT by adding micro-shift at the predicted position and bounding box adaptation.