Residual-Based Learning Efficient Transformer Network for Maneuvering Target Tracking

We investigate the challenging problem of maneuvering target tracking (MTT) without relying on model prior knowledge. Model-based approaches typically combine several simple models to approximate the state evolution of a real target, which may suffer from model mismatch issues under complex kinetic patterns. To address this, we propose a novel data-driven MTT approach by leveraging transformer network. This approach exploits prediction residuals as network inputs rather than original measurements to learn target maneuvering features, enabling rapid convergence on real-world datasets. Numerical experiments demonstrate our proposed approach's tracking accuracy and computational efficiency compared to the traditional model-based interacting multiple model (IMM) approach in various tracking cases. Codes available at https://github.com/poem-flower/RBMTT.git