Answers are grounded in this paper and cite their sources.
Unmanned surface vehicle (USV) localization and tracking in complex marine environments are easily affected by model mismatch, uncertain noise statistics, and abnormal observations, which may degrade the estimation performance of the conventional extended Kalman filter (EKF). To improve the adaptability of recursive filtering under uncertain conditions, this paper proposes a network-aided robust EKF estimation framework, termed REKFNet. While preserving the prediction–update structure of the EKF, the proposed method introduces a GRU–attention module to learn uncertainty-aware temporal features from historical residual information, thereby assisting the adaptive adjustment of the Kalman gain. Meanwhile, an anomaly-aware processing mechanism based on an LSTM autoencoder is designed before the measurement update to reduce the influence of contaminated observations on posterior estimation. Experimental results show that, compared with conventional filtering methods and representative learning-based baselines, REKFNet achieves lower estimation errors under model mismatch, noise uncertainty, and abnormal observations, and exhibits better trajectory consistency on real-world trajectory data. Sparse multi-target experiments further suggest the potential applicability of the proposed estimator when combined with a lightweight association strategy in low-density shore-based tracking scenarios.
No takes yet. Share an insight, caveat, or question.
Yan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: