Cross-Attention Transformer for Coherent Detection in Radar Under Low-SNR Conditions

Detecting weak echoes from low-RCS targets in pulsed radar systems presents significant challenges, as conventional coherent accumulation methods require extended dwell times that reduce data rates and suffer from target-motion-induced migration. We propose RD-Transformer, an end-to-end attention-based architecture that reformulates coherent integration as a learned feature fusion problem. The framework integrates multi-pulse transpose preprocessing, dual-path self-attention encoders for transmitted and received signals, and a cross-attention decoder to extract transmit-receive interaction features. A tunable sigmoid-based gating mechanism enables flexible false alarm control during inference. Experiments on synthetic pulsed-radar data demonstrate that, under identical false alarm constraints (P_fa = 1 × 10^-2 to 1 × 10^-5) and using only 10 coherent pulses, RD-Transformer reduces the required SNR by 14-20 dB compared to optimal energy detection across Swerling I-IV target fluctuation models, validating the effectiveness of learned coherent accumulation for weak target detection.