A computationally efficient distributed Bayesian filter with random finite set observations

This paper presents a distributed Bayesian filter for tracking a target in the presence of random missed detections and false alarms using a sensor network, in which the Bayesian recursion as well as our proposed multi-sensor posterior fusion is carried out via Gaussian mixture (GM). For better communication and computation efficiencies, only properly selected Gaussian components in the GMs are disseminated and fused between neighbor sensors in which the components are selected following the principle of principal component analysis. Linear/arithmetic average fusion is realized for which thresholds used in GM merging and pruning operations are theoretically derived by the Occam's window method. Furthermore, an improved distributed flooding protocol is devised for GM communication over the network which enables parallelization of internode communication and fusion operations and reduces the node-memory cost. It is demonstrated that under reasonable assumptions, it yields the same result as the original flooding algorithm while having lower node-memory requirements. Our proposed approach is compared with the state-of-the-art approach in both simulation and experiment scenarios for target tracking.