Learning Multisensor Confidence Using a Reward-and-Punishment Mechanism

In many application scenarios, multiple sensors are deployed in an observation area to per-form various monitoring tasks. The observations of the participating sensors are fused together to obtain a more accurate and improved decision about the occurrence of an event. However, the sensors deployed in an environment do not have the same confidence level due to their differences in capabilities and imprecision in sensing and processing. The confidence in a sensor represents the level of accuracy in accomplishing a task that can be computed either by comparing the current observation with a reference data set or by performing a physical investigation-both of which are not feasible in a real scenario. Nevertheless, it is essential to know how the sensors are performing with respect to the objective tasks. This paper addresses this issue and proposes a novel reward-and-punishment mechanism to dynamically compute the confidence in sensors by leveraging the differences of the individual sensor's opinion. Experimental results show the suitability of utilizing the dynamically computed confidence as an alternative to the accuracy of the sensors.