Tracing-KalmanNet: Deep Kalman Filtering for Nonlinear and Intermittent Inertial Data

Kalman filters are fundamental to attitude estimation and inertial navigation, effectively mitigating sensor errors and drift to enhance the accuracy and stability of systems. However, their performance degrades significantly in complex scenarios characterized by nonlinear dynamics, intermittent observations, and unknown noise statistics. To address these challenges, we propose Tracing-KalmanNet (TKN), a deep learning-enhanced filtering framework that integrates neural networks into the Kalman gain computation. Unlike traditional methods, TKN operates without prior knowledge of noise distributions or covariance models, achieving robust performance in highly nonlinear and data-sparse environments. Moreover, by leveraging the temporal context of historical inertial data, TKN dynamically adapts its parameters to improve generalization and resilience to model mismatch. Extensive experiments demonstrate that TKN consistently outperforms classical and learned Kalman filter variants under discontinuous and nonlinear inertial measurements, offering a promising solution for attitude estimation in degraded sensing conditions.