• 
• 
• 
• 
• 
• 

Ensemble Tracking

We have developed a general tracker that can be used in various scenarios such as surveillance and human-computer interaction. The tracker can track a wide variety of objects (e.g., pedestrians, faces, cars, boats) using either a static or dynamic camera. It is computationally efficient, robust, runs in real time and can work with different image types (e.g., color, gray-scale, infra-red).

Background & Objective:  Tracking algorithms usually focus on the object to be tracked, while neglecting the environment in which the object is moving. Furthermore, most tracking algorithms work with pre-defined features that might not be able to distinguish the object from the background. Our tracker can automatically find the best set of features to separate the object from the background, as well as integrate and update this information over time.

Technical Discussion:  We treat tracking as a binary classification problem where the object is to be distinguished from the background. We use AdaBoost to construct, online, an ensemble of weak classifiers that are trained one per frame. In addition, we developed an online update rule for ensemble training that allows us to deal with time-varying distributions that arise from the change in appearance of the object and the background. The result is a dynamic ensemble of weak classifier that allow for robust and computationally efficient tracking.

Technology Area:  Computer Vision

Modification Date:  September 14, 2007