Fusing sensors with uncertain detection performance

Abstract

Sensor fusion is the notion of combining the data from two or more sensors in order to obtain enhanced performance compared with that of the individual sensors. In addition, Signal Detection Theory can be used to monitor how well a sensor operates. That is, through the number of hits, misses, false alarms and correct rejections a sensor registers, we gain a better understanding as to how reliably it performs. Typically, the performance of a sensor is given in terms of its probability of detection and probability of false alarm, which may not be well characterised. In this paper, we use the Transferable Belief Model to fuse two sensors where there is uncertainty in their performance, so that if two sensors give a report, for example, we can estimate the likelihood of the target being present. We also show that when we have known prior probabilities our result is equivalent to the Bayesian case. A numerical example, as well as entropy measures, are also discussed.

A Robust Fusion Algorithm for Sensor Failure

“…When , the case where no sensors fail, the rule simplifies to the traditional fusion rule defined in (2). When…”

“…It has been shown in [1] that Bayes Risk Error is a Bregman divergence between the actual and estimated prior distribution probabilities. Davey uses a Transferable Belief Model, an inspiration from Dempster-Shafer Theory, to successfully manage uncertainty in the prior distribution of the target variable [2]. Our work here, however, focuses on uncertainty in the likelihood itself due to sensor failures.…”

Analysis and Comparison of Probability Transformations for Fusing Sensors with Uncertain Detection Performance

“…In [4], the Transferable Belief Model (TBM) [6,7], a variant of DST, was used to fuse two sensors at the decision level (is there a target or not). It was found that uncertainty in the prior probability of a target being present lead to uncertainty in the fused belief mass.…”

“…It was found that uncertainty in the prior probability of a target being present lead to uncertainty in the fused belief mass. In [4] this was resolved by using the pignistic transform. However, alternative methods for transforming uncertain belief functions into probabilities have been proposed.…”

“…However, alternative methods for transforming uncertain belief functions into probabilities have been proposed. Here, we extend the work of [4] and analyse several probability transformations that satisfy the minimal criterion of [5].…”