Adaptive Deep Kalman Filtering for TDOA Maneuvering Target Localization

Recent research emphasizes passive localization using time difference of arrival (TDOA)-based methods, valued for their concealment, accuracy, and robustness. Kalman filter and its variants are capable to solve TDOA tracking problem, utilizing temporal information and they show great performance in high-noise scenarios. However, In maneuvering target tracking, conventional Kalman filters face challenges due to their assumptions of linearity, Gaussian noise, and difficulties in modeling highly dynamic targets. While some Kalman filter variants provides adaptability, it may still struggle with rapid maneuvers. To address these challenges, this paper proposes the adaptive deep Kalman filter (ADKF), as an innovative state estimator designed to improve Kalman filtering under maneuver dynamics with incomplete information. ADKF utilize linear recurrent unit (LRU) to learn the residual covariance matrix. It achieves enhanced stability in the filtering process, mitigating the limitations observed in other approaches. This adaptive learning mechanism empowers ADKF to improve tracking accuracy in the face of target maneuvers, providing a more robust and versatile solution for maneuvering target tracking. Furthermore, ADKF consistently outperforms classic filtering methods across various maneuver levels and noise conditions.