Modified Swin Transformer-Based Continuous-Discrete Extended Kalman Filtering for Bearings-Only Tracking With Unknown Noise Statistics

To overcome the accuracy degradation for bearings-only tracking (BOT) under time-varying, unknown process noise, the adaptive square-root continuous-discrete extended Kalman filter (CD-EKF) algorithm enhanced by a modified Swin Transformer is proposed. First, the square-root CD-EKF framework is employed to mitigate the adverse effects of limited sensor observation. Second, a modified Swin Transformer model is developed to extract statistical features of the process noise and adaptively estimate the corresponding covariance matrix. Furthermore, the shifted window-based self-attention mechanism makes the Swin Transformer capable to extract the hidden statistical information in the state without additional constraints. In addition, a repetition module is introduced to help the Swin-adaptive Kalman filter (AKF) in understanding the statistical relationship between position and velocity state. The results of the numerical experiment using BOT examples in constant-velocity and constant-turn scenarios, and the results of the experiment on the Oxford RobotCar dataset demonstrate that the proposed adaptive filter exhibits superior robustness and accuracy compared to the existing filters in addressing measurement insufficiency and process noise uncertainties.