Generalized Choquet fuzzy integral fusion

Sensor fusion plays an important role in many application domains. No single source of information (decision or feature) can provide the absolute solution when detection and recognition problems become more complex and computationally expensive (e.g., in land mine detection). However, complementary information can be derived from multiple sources. In this paper, we build a decision-based fusion system based on the uncertainty approach utilizing an extension of the Choquet fuzzy integral (generalized Choquet fuzzy integral, GCFI). The difference between the standard Choquet fuzzy integral and the GCFI is that the GCFI integrates vectors of fuzzy numbers instead of vectors of numeric membership values. The system is applied to a land mine detection problem. The fuzzy vectors represent uncertainty in both the confidence and location estimates of several detection algorithm outputs. The results show a huge improvement in the probability of detection and a reduction in the false alarm rate over the best algorithm and two numeric fusion schemes, i.e., the average confidence and a decision level fusion with the numeric Choquet fuzzy integral. The GCFI obtains 100% probability of detection at 0.02 false alarm rate per square meter on a large test set, whereas the best detection algorithm and the average confidence achieve only 91% and 96% probability of detection at that rate. Additionally, at 0.02 false alarm rate, decision level fusion with the numeric Choquet fuzzy integral reaches only 87% probability of detection.