Automated selection of fusion parameters through a segmentation of multi-sensor ROC curves

This research builds upon a mathematically proven optimal decision fusion technique exploiting the Neyman-Pearson (N-P) test. The algorithm requires three parameters for each sensor input, so the number of fusion parameters increases linearly. A new method presented in this paper, relies on two meaningful external parameters defined by an operator. They trigger an automated selection of the remaining internal parameters for all sensory inputs. The outcome is a set of quasi-optimal parameters. The method exploits a segmentation and alignment of individual ROC curves into similar regions of compatible confidence levels. Experimental results are shown for synthetic test data and real-world mine hunting data. This new method allows for an automated dynamic integration of system components into a system, gives consistently better performance, and requires two parameters only regardless of the number of sensor inputs.