均值融合跟踪概念、理论与方法

The distributed multi-sensor fusion tracking problem lies at the forefront of research in the field of information fusion. The core of the research is the fusion of local posteriors (mainly in the form of posterior densities or their statistical moments) obtained by different sensors to obtain more accurate, reliable, and consistent target state estimates. The emerging averaging approach that fuses the probabilistic densities has received considerable attention in recent years, given in two distinctive forms, namely arithmetic and geometric average fusion. It is able to efficiently deal with the unknown correlation between sensors and has demonstrated superior performance for both single-target and multi-target tracker designs, achieving internode consensus approximately or asymptotically. This paper systematically reviews the Bayesian filter tracking model, random finite set theories, and the concepts of variable fusion and density average fusion and provides a comparative analysis of the statistics and informatics properties of the two density average fusion methods. The analysis encompasses covariance consistency, mean square error, mode-preservation capacity, suboptimality, etc. Furthermore, the paper explores various strategies for designing fusion weights and presents analytical solutions for averaging fusion under representative random finite set distributions. A unified framework based on probability hypothesis density averaging fusion, capable of deriving the average fusion results for various random finite set filters, is highlighted, along with a comparison of the strengths and limitations of the two average fusion methods. In conclusion, in response to the various realistic application needs, the remaining challenges and future research trends of average density fusion regarding complex targets (such as resolvable targets and extended targets), limited sensors (such as limited communication bandwidth, fusion delay, dynamic/limited sensing range), sensor network attacks and privacy protection, and sensor resource management are envisioned.