Heterogeneous multi-sensor fusion for PHD filter in decentralized sensor networks

Abstract

This paper addresses distributed multi-sensor multi-object tracking based on probability hypothesis density (PHD) filter. Due to the scalar fusion weights, existing distributed fusion methods only work well in cases where the information confidence within multi-object densities (MODs) returned by local filters remains unchanged over the object state space. Therefore, severe performance degradation can be observed as, in practice, the information confidence of an MOD is usually space-varying and can be utterly different due to the limited sensor abilities and various environment impacts. We make three contributions towards addressing this problem. Firstly, we propose a heterogeneous fusion method for multi-object Poisson process (MPP) MODs. The resulting MOD has a parallelizable structure that allows multiple clusters of sub-densities to be fused independently based on different sets of heterogeneous fusion weights. Secondly, a density-based heterogeneous fusion weights design method is proposed, in which the fusion weights are determined according to the relative information confidence of MODs. Lastly, to enhance the adaptability and robustness of the proposed fusion method, we extend it to decentralized sensor networks in conjunction with consensus and flooding protocols, respectively. The implementation of the proposed fusion algorithms using Gaussian approximation is also presented. The efficacy of the proposed fusion methods is demonstrated in various numerical experiments, including a challenging scenario in which each sensor in a decentralized network has dynamic detection probability and space-varying detection accuracy. In the experiments, the proposed method significantly outperforms the state-of-the-art in distributed fusion. In fact, the Optimal Sub-Pattern Assignment (OSPA) distances returned by the proposed method are at least 38.8% less than other solutions.