Abstract
Tracking weak targets in low signal-to-noise ratio (SNR) radar environments remains challenging for conventional detect-before-track (DBT) methods, since early detection may discard useful information. This letter proposes a Transformer based particle filter track-before-detect (PF-TBD) framework. First, we construct a measurement-driven particle likelihood from unthresholded 4D-FFT data by Gaussian-weighted coherent local aggregation and a GLRT-based energy-ratio normalization, achieving robust tracking without signal amplitude priors while reducing sensitivity to local motion-prediction errors. Second, inspired by self-attention imputation for time series (SAITS), we develop a trajectory imputation module which utilizes diagonal causal masked self-attention trained in a SAITS-style masked self-supervised manner on reliable PF-TBD trajectory estimates for recovering missing trajectory segments during low-confidence intervals. Then, during low-confidence intervals, the imputed states are fed back into the TBD particle filter in a recovery mode to maintain recursive information accumulation. Simulations show that DBT baselines fail below 5 dB SNR, while the proposed method maintains sub-meter accuracy down to 0 dB and outperforms the conventional PF-TBD across all tested SNR levels. Under consecutive 4-frame detection gaps, closed loop imputation reduces position error by 14.2%, validating the effectiveness of trajectory recovery.