31 December 2024 in IEEE Sensors Journal
To achieve optimal multi-target fusion tracking using an active sonar’s kinematic, spread angle measurements, and a passive sonar’s bearings-only measurements, a federated fusion algorithm of Gaussian mixture probability hypothesis density (GM-PHD) filters is proposed. Our algorithm features a hierarchical structure consisting of a fusion node and two sonar nodes. Based on a newly defined target state, the observation model of spread angle measurements is established as the premise of fusing this data. Two local filters of sonar nodes transform heterogeneous measurements into isomorphic state estimates, erasing data heterogeneity. A master filter of the fusion node can supplement missed detections. Multi-target density fusion is decomposed into multiple single-target estimate fusion by estimate association. We derive the optimal single-target estimate fusion methods in the absence and presence of missed detections. The covariance upper-bounding technique is introduced to eliminate inevitable correlations. Simulations demonstrate our algorithm outperforms the existing centralized and distributed heterogeneous fusion tracking algorithms, and the relative weights of the master and local filters can be flexibly adjusted.
3 March 2025 in IEEE Sensors Journal
To achieve the optimal uncertain multi-target decentralized tracking fusion, a federated density fusion algorithm of Gaussian mixture-probability hypothesis density (GM-PHD) filters is proposed. Our fusion algorithm requires a hierarchical structure. Each sensor node runs a local GM-PHD filter to extract multi-target state estimates from sensor measurements. Besides running a master filter that only predicts the fusion result of the previous moment, the fusion node also collects, associates and merges the GM-PHDs of the master and local filters after eliminating the correlations among all filters. The association results can be judged as survival target detection, missed detection, target new-birth or target vanishing. Each association situation employs a different optimal merging method. To eliminate correlations resulting from feedback and common process noise, the covariance upper-bounding technique is introduced. This technique returns the cross-covariances among all filters to zeros by increasing their auto-covariances and the process noise covariance. It is proven that using the master filter, feedback and covariance upper-bounding technique, the fusion node can obtain the same optimal accuracy as the centralized fusion. Simulations demonstrate our algorithm, with appropriate parameters, can obtain the optimal tracking accuracy and outperforms the existing arithmetic average (AA) fusion algorithm.
1 June 2024 in Digital Signal Processing
24 March 2025 in IEEE Signal Processing Letters
We achieve the optimal Arithmetic Average (AA) fusion algorithm of Gaussian Mixture Probability Hypothesis Density (GM-PHD) filters in a hierarchical structure. First, the optimal single-target estimate fusion is derived, during which the prior estimate is indispensable. Then, the derived optimal estimate fusion is employed as the merging method of the AA fusion. A master filter dedicated to computing prior density is introduced, so our fusion algorithm features a hierarchical structure. Experiment results evidence our algorithm's optimality and superiority over the standard AA fusion.
20 February 2025 in Digital Signal Processing
20 May 2025 in Signal Processing