Computer Science · Artificial Intelligence
A new approach for evaluating the classification performance of multi-sensor fusion systems
See on the mapAbstract
We present a new approach for evaluating the classification performance of multi-sensor fusion systems. A common problem in target tracking is to use one/more sensors to observe repeated measurements of a target's features/attributes, and in turn update the targets' posterior classification probabilities. This paper introduces new metrics and approaches to quantify the performance of a single/multi-sensor classification system. We show minimal conditions under which sensor(s) will classify all targets perfectly. We also derive exact and approximate formulas for efficient calculation of the long-run classification performance, in a manner analogous to the use of the Kalman filter for kinematic performance. We also present a methodology to evaluate the performance of a classification system with sensors of varying quality.