A Target Tracking Model for Weak Radar Targets Based on Track-Before-Detect

Radar target tracking under low signal-to-noise ratio (low-SNR) conditions remains a critical challenge. For targets such as stealth aircraft and low-altitude unmanned aerial vehicles (UAVs), radar echoes are often extremely weak and submerged in noise, rendering conventional detection and tracking ineffective. To address this issue, we propose a method that combines a Convolutional Long Short-Term Memory (ConvLSTM) network, U-Net, and the dynamic-programming track-before-detect (DP-TBD) algorithm. Specifically, DP-TBD generates multi-frame accumulation maps, which are fed into the network. U-Net extracts spatial features from each accumulation map, ConvLSTM models temporal dependencies over the accumulation-map sequence, and a Squeeze-and-Excitation (SE) module suppresses noise-dominated channels. The network outputs a final-frame confidence map for thresholding and track extraction. Experiments show that the proposed method improves tracking performance in low-SNR environments, achieving a 90\% tracking rate with a false alarm rate below $10^{-5}$ at an RD-map SNR of 3~dB.