Reducing Drift in Parametric Motion Tracking

Ali Rahimi, Louis-Philippe Morency, Trevor Darrell

There is a technique for reducing drift in vision-based trackers. Louis-Philippe Morency and Ali Rahimi have been working on this since 2000.

Modern techniques for tracking the location of objects range from GPS and magnetic field sensing, to sonar and vision-based techniques. Traditionally, most of these techniques are either inaccurate, limited to a few domains, or suffer from drift. Accurately tracking an object for prolonged periods of time is difficult no matter what technology is used.

Most trackers use a Markov chain as their measurement model, and use a Kalman filter to compute optimal pose estimates. This research is about building an appearance model of the target while we track it. This appearance model is refined during tracking and helps the track do its job.

The technique has been evolving since 2001. Here is info about an older version of the tracker.