Most high level vision tasks (behavior analysis, action recognition, etc.), especially in the context of video surveillance, often focus on targets of interest within the given scene. Therefore, a key task that needs to precede these activities is the identification of important targets. In this work, we focus on targets of interest as being those individuals that persist across the entire input video sequence, since many applications in surveillance naturally arise from this. More specifically, we address the problem of automatic localization and tracking of persistent targets in video sequences and study the problem in the domain of multiple pan-tilt-zoom (PTZ) cameras.
We propose a novel Multiple Instance Learning (MIL) framework which consists of a softmax-based combining function along with a logistic model for the instances employing log covariance features. The learned target models can be updated in an online manner and can also be used to learn models for multiple targets present in the scene, and track them in a “tracking-by-detection” mould. We develop active camera control and coordination infrastructure to extend the proposed approach to multiple cameras. We also investigate a one-class version of the Multiple Instance problem and develop a support-vector machine based algorithm based on prior probabilities on instance labels. Lastly, we perform detailed experiments to demonstrate the validity and usefulness of the proposed approach to localize targets in various scenes using commercial-grade surveillance cameras.