A Robust Adaptive Sliding-Window Variational Bayesian Filter with Unknown Noise Covariance Statistics
Abstract Target tracking performance is severely limited under unknown noise covariance matrices, as conventional filtering algorithms fail to produce satisfactory optimal state estimates. To address this bottleneck, this work develops a novel adaptive robust sliding‑window variational Bayesian filter. Within this framework, both measurement noise covariance matrix (MNCM) and process noise covariance matrix (PNCM) are modeled with inverse Wishart (IW) distributions. The variational Bayesian (VB) is adopted to jointly iterate over target state, MNCM and PNCM for simultaneous estimation. The obtained optimal estimations are fed as time varying inputs into multiple fading factors to compensate the prediction error covariance matrix (PECM), which boosts overall estimation precision. Besides, an enhanced adaptive sliding‑window strategy is constructed to calibrate prior states from posterior knowledge, yielding better estimation accuracy together with higher computational efficiency. A square root implementation is further derived to guarantee numerical stability. It maintains symmetry and positive definiteness of covariance matrices and avoids filter divergence. Simulation experiments verify that the presented approach can effectively mitigate performance degradation induced by imprecise noise covariance, and outperforms mainstream existing filters in target state estimation.
Authors
- Minzhe Li (ORCID: https://orcid.org/0000-0002-7168-6949)
- Peng Gu (ORCID: https://orcid.org/0000-0002-7155-3388)
- Zhongliang Jing (ORCID: https://orcid.org/0000-0003-1759-8785)
Institutions
- Shanghai Jiao Tong University (CN)
- Jiangsu Vocational College of Information Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-11498837/v1
- Primary Topic
- Target Tracking and Data Fusion in Sensor Networks
- Type
- preprint