Adaptive Hybrid Bayesian-Deep Learning Framework for Maneuvering Target Tracking in Low-SNR Radar Environments

Tracking highly maneuvering targets in low Signal-to-Noise Ratio (SNR) environments remains a critical challenge in radar signal processing. Traditional Bayesian filters (e.g., EKF, IMM) fail during abrupt target maneuvers, while modern Deep Learning (DL) models (e.g., LSTM) lack robustness in low-SNR and data-scarce scenarios. This paper addresses this gap by proposing a novel Adaptive Hybrid Deep-Learning Framework (AHDF). The AHDF integrates an Extended Kalman Filter (EKF) core, which provides physical model stability, with a parallel Long Short-Term Memory (LSTM) network trained to estimate non-linear motion patterns by observing the filter's residuals. The core innovation is an Adaptive Fusion mechanism, implemented via a Deep Reinforcement Learning (DRL) agent, which dynamically learns the optimal policy for adjusting the EKF's process noise covariance (Q) based on the combined outputs of the EKF and LSTM. We conduct extensive Monte Carlo simulations across various maneuvering scenarios and SNR levels. Results demonstrate that the proposed AHDF significantly outperforms baseline EKF, IMM, and pure LSTM trackers, reducing tracking RMSE by up to 47.5% compared to the IMM during maneuvers, and decreasing the track loss rate by 95% compared to the pure LSTM in low-SNR conditions. The AHDF demonstrates a 3x faster convergence time post-maneuver, proving the efficacy of the DRL-based optimal fusion policy.

کلیدواژه ها:

Adaptive Filtering ، Deep Learning ، Deep Reinforcement Learning (DRL) ، Hybrid Filter ، Kalman Filter Target Tracking