An End-to-end Instance Segmentation Network for Radar Target Detection and Association

Radar systems operating in complex environments face challenges such as dense target distribution, trajectory overlap, and fluctuating scattering intensity. Traditional threshold and filter-based methods often fail to achieve both high detection accuracy and reliable target association. An end-to-end instance segmentation network is proposed to jointly address target detection and association. Pulse-compressed signals are converted into time-range images, and an encoder-decoder network is employed to perform pixel-level detection and feature embedding. Background constraints and embedding losses are incorporated during training to improve feature separability, while unsupervised clustering in the embedding space enables automatic target distinction and association during inference. Experimental results show that the proposed method achieves superior accuracy, robustness, and efficiency compared with conventional approaches, providing a practical framework for intelligent radar detection and track initiation.