Sensor Fusion

Program Description

The complexity and unpredictability of the large scale search and surveillance problem requires the use of revolutionary new methodologies in a formal mathematical framework. Our work focuses on several methodologies, including particle filtering and Gaussian mixture models. Our work also focuses on demonstrations, such as in the DARPA Grand Challenge.

One specific approach is a novel computational uncertainty management methodology which exploits symmetries between estimation and planning: Populations of candidate models and candidate actions continuously co-adapt while targeting weaknesses in each other, yielding a set of robust models and robust actions to improve models. One key direction in our research is developing a decentralized estimation/control by co-evolving alternative candidate models (that match fused data, make predictions) and candidate vehicle actions (that generate disagreement between models). A second key direction is developing the mathematical foundations of integrating decentralized particle filters (for sensor fusion) and Gaussian mixture models (for non-Gaussian uncertainties, data exchange) in order to develop accurate non-Gaussian estimates and use in higher level planning. The program focuses on decentralization of all elements, mathematical foundations, and real time implementation.

Sponsor

Lockheed Martin, NASA, NSF

Staff and Students

Publications

• J. Sullivan, M. Campbell, and H. Lipson, “Particle Filters as Exploration Tools for Autonomous Rovers,” AIAA Guidance, Navigation and Control Conference, Aug 2005.
• Y. Fang, M. Campbell, “Probability Map Building Algorithms Design for an Unknown Dynamic Environment,” International Conference on Natural Computation, 2005.