Projective method for generic sensor fusion problem

In a multiple sensor system, each sensor produces an output which is related to the desired feature according to a certain probability distribution. We propose a fuser that combines the sensor outputs to more accurately predict the desired feature. The fuser utilizes the lower envelope of regression curves of sensors to project the sensor with the least error at each point of the feature space. This fuser is optimal among all projective fusers and also satisfies the isolation property that ensures a performance at least as good as the best sensor. In the case the sensor distributions are not known, we show that a consistent estimator of this fuser can be computed entirely based on a training sample. Compared to linear fusers, the projective fusers provide a complementary performance. We propose two classes of metafusers that utilize both linear and projective fusers to perform at least as good as the best sensor as well as the best fuser