A Screened IMM-GRU Method for Low-Probability-of-Intercept Maneuvering Target Tracking

To address the difficulty of accurate motion-model prespecification and delayed model switching in maneuvering target tracking, this paper proposes a joint model- and network-driven screened Interacting Multiple Model-Gated Recurrent Unit (IMM-GRU) method. In the filtering stage, multiple network outputs are screened using a smooth local reference trajectory constructed from measurements, providing a physically meaningful criterion for rejecting inconsistent estimates while retaining candidates compatible with local track evolution. For radar low-probability-of-intercept (LPI) operation, a difference-enhanced attention Bidirectional GRU (Bi-GRU) classifier identifies target motion modes during radar-silent intervals and supplies model priors for predictor screening. A screened IMM-GRU predictor then selects candidates using model probability, prediction uncertainty, inter-network disagreement, and local kinematic consistency. Cumulative prediction variance and measurement consistency are used to schedule radar radiation. Simulation results show that the proposed filtering and prediction modules improve tracking and extrapolation accuracy. Under the equal-accuracy LPI criterion, the proposed framework achieves the lowest four-scenario average radiation ratio at every prescribed MRMSE level from 40 to 250 m and is the only method that satisfies the 30 m MRMSE requirement in all four scenarios. These results demonstrate improved maneuvering-target tracking within the examined model and noise ranges while reducing unnecessary radar exposure.