Data-Driven Maneuvering Target Tracking Model of Attention-based Gated Recurrent Unit and Adaptive Unscented Kalman Filter

Target tracking is a key technology for achieving situational awareness. A data-driven maneuvering target tracking model based on an encoder-decoder structure is proposed to improve tracking precision under challenging conditions such as high speed, strong maneuverability, and non-Gaussian noise. The encoder employs an attention-based Gated Recurrent Unit (attention-GRU) to capture motion state and temporal dependencies. The decoder utilizes an adaptive Unscented Kalman Filter (UKF) optimized by Expectation-Maximization (EM), which learns the noise distribution characteristics of the data and dynamically estimates UKF parameters. The target state estimation is achieved through the adaptive UKF. Experimental results show the proposed model effectively tracks high-speed maneuvering targets in simulation, including hypersonic ones. The proposed model significantly outperforms KF, UKF, Long Short-term Memory (LSTM)-KF, and LSTM-UKF in reducing the Root Mean Square Error. Additionally, the tracking precision for ground-based radar detection under glint noise has demonstrated the robustness of the model.