Distributed heterogenous multi-sensor fusion based on labeled random finite set for multi-target state estimation using information geometry

Abstract

This paper considers the challenging problem of heterogenous multi-sensor fusion using information geometry theory. A novel distributed heterogenous fusion method with labeled random finite set (RFS) multi-target densities is proposed to address multi-sensor multi-target tracking problem. The Fisher information distance (FID) is used to characterize the information acquisition and sensing capability of distributed sensor nodes by analyzing measurement model of sensor nodes via geodesic computation on statistical manifold. A scalar fusion weight is assigned to each local multi-target density in order to characterize its relative information confidence under the different architecture of sensor nodes. The fusion weights are adjusted according to the contribution of each sensor node to global multi-target density. Thus, the proposed distributed heterogenous fusion serves as an adaptive algorithm for multi-sensor multi-target tracking within the monitored area. Aiming to achieve multi-target tracking under the limited sensing capabilities of sensor nodes, we firstly formulate the general framework of distributed heterogenous fusion strategies based on sensor information accumulation. Secondly, we present both numerical and approximate solutions to computing the geodesic on the manifold for FID calculations, furtherly for determining locally tailored fusion weights for different sensor nodes. Finally, we present several simulation results to validate the efficacy of our proposed distributed heterogenous fusion algorithm for multi-target tracking. These simulations involve three typical types of sensor nodes where each sensor agent performs a multi-scan generalized labeled multi-Bernoulli (GLMB) smoothing algorithm, demonstrating superior performance in multi-target state estimation.