Learnable geometric method on multi-sensor weighted evidence fusion

Focused on the shortcomings of dealing with the evidences conflict in Dempster-Shafter combination rule and other improved algorithms, a new learnable geometric method on weighted data fusion is presented. In this method, a linear space is spanned by basic probability assignment vector over discernment frame, and the historical evidence data of each sensor in a period of time are mapped into the space to form the geometric model of evidence point distribution. Then the credibility of each sensor is evaluated by credible radius of evidence set of the sensor, and the optimal weighted allotment of different evidence sources is acquired. Along with the increasing of new point in evidence set, the credibility of sensor tends to an accurate and steady value. Because the conflict and consistency can be characterized by the Euclidean distance of evidences in the linear space to describe the rejection or support of each focus element, the weighted combination rule can be expressed and calculated conveniently by the method presented. Experiment results show that the method can estimate the credibility of sensor accurately and improve the combination rule effectively.