Particle filters for real-time object tracking

Visual tracking of moving objects from a moving camera in the presence of background clutter is an active area of research in computer vision. Recently particle filters or the condensation algorithm [IsardBlake98] have shown to be very suitable to perform real-time tracking in cluttered environments. Unlike Kalman filters, which are limited to gaussian probability distributions, particle filters are able to respresent multimodal distributions. A discrete set of samples or particles represents the object-state and evolves over time driven by the means of "survival of the fittest". Nonlinear motion models can be used to predict object-states.

In this project particle filters are investigated and extended in the field of real-time tracking in robot vision. Research focuses on the following topics:
  • Multi object tracking: Clustering algorithms are used to identify different objects so that subpopulations are able to evolve independently. Various techniques to deal with overlapping, entering and leaving objects are discovered.
  • Dynamic models: Beside using random walk dynamics or learned models it is investigated how far self-adaptive dynamic models are able to cope with sudden changes in motion.
  • Machine learning of object models: A variant of the adaboost learning mechanism is used to learn feature-based classifiers from training sets (see [ViolaJones01]). These classifiers are used to model observation probabilities in particle filters.
Tracking a face using learned features
Tracking a colored ball

References

  • [IsardBlake98]M. Isard and A. Blake , "Condensation - Conditional Density Propagation for Visual Tracking, " International Journal of Computer Vision, vol. 29, no. 1, pp. 5-28, 1998.
  • [ViolaJones01]P. Viola and M.J. Jones, "Robust Real-time Object Detection," Tech. Rep. CRL 2001/01, Compaq Cambridge Research Laboratory, Cambridge MA, 2001.

Contact

André Treptow, Tel.: (07071) 29-778970, treptow@informatik.uni-tuebingen.de