3D Radar Target Prediction and Tracking Using DAST-Network

In 3D radar target tracking, nonlinear motion and complex noise-induced accuracy degradation remain formidable challenges. This paper introduces a deep learning model combining dynamic attention, gated recurrent units (GRU), and Inception modules. The model utilizes the multi-scale feature fusion capability of the Inception module to extract both local and global spatial features of radar point clouds, and employs the GRU network to model the temporal dependencies of the target’s motion state. Furthermore, a hybrid motion prior is developed by fusing the 3D constant velocity (CV) and constant acceleration (CA) models, enabling superior adaptability to targets undergoing uniform motion or acceleration. Simulation experiments under various noise conditions show that the proposed model reduces the root mean square error (RMSE) of position estimation by approximately 51.5% to 67.7% compared to traditional Kalman filter (KF), extended Kalman filter (EKF), and unscented Kalman filter (UKF) methods. This research provides a new solution for robust 3D radar target tracking in complex environments.