Filtering and fusing compass and gyrometer data using Guess filter

Most papers dealing with data fusion try to use a single tool to provide the "best" estimation. But, as pointed out by Prade-Dubois (1988), most tools dealing with imperfect information aim at different purposes. In the previous paper (Strauss et al. (1996)), the Guess filter has been presented. This filter aims at using three different error theories together to obtain an estimation that combines robustness, accuracy, reliability and easy setup. The possibility theory handles precision, the statistical theory is used to reduce uncertainty, and the rough set theory allows a robust and easy computation of the resulting filter. In the Guess filter, data are represented as a possibility distribution, and because of this representation, data issued from different sensors can be combined at both high and low level. Fusion at a low-level takes advantage of redundancy to reduce the overall uncertainty and thus to increase accuracy. Fusion at a high-level reduces the influence of inadequacy in data modeling. This method has been implemented on a submarine robot. Experimental results are presented.