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

Area of Science:

• Radar Systems Engineering
• Signal Processing
• Machine Learning

Background:

• Detecting low-RCS targets in pulsed radar is challenging due to conventional methods requiring long dwell times.
• Target motion causes migration, complicating coherent accumulation and reducing data rates.

Purpose of the Study:

• To propose an end-to-end attention-based architecture, RD-Transformer, for improved weak target detection.
• To reformulate coherent integration as a learned feature fusion problem.

Main Methods:

• Developed RD-Transformer, an architecture integrating multi-pulse transpose preprocessing and dual-path self-attention encoders.
• Employed a cross-attention decoder for transmit-receive interaction features and a sigmoid-based gating mechanism for false alarm control.

Main Results:

• RD-Transformer reduced required SNR by 14-20 dB compared to optimal energy detection across Swerling I-IV models.
• Achieved this improvement using only 10 coherent pulses under stringent false alarm constraints (1x10^-2 to 1x10^-5).

Conclusions:

• Learned coherent accumulation via RD-Transformer is effective for weak target detection in pulsed radar.
• The proposed method offers a significant advancement over conventional techniques, enhancing radar sensitivity and data rates.