Design of a Hybrid OS-CFAR Network for Improving Radar Target Detection and Tracking Performance

This study proposes the Hybrid Ordered Statistic Constant False Alarm Rate (OS-CFAR) network that replaces the sorting- based order statistic computation with a neural estimator while preserving the scale-factor-based decision rule. OS-CFAR is extensively used in airborne radar for its robustness to non-stationary clutter, but its per cell-under-test sorting of the reference window dominates the computational cost, limiting real-time radar applications. To this end, the proposed network combines a U-Net encoder–decoder with a convolutional block attention module to estimate the detection threshold over the entire Range-Doppler map in a single forward pass, while a NeuralSort-based auxiliary loss imposes the order statistic constraint in a differentiable form during training. Because the explicit scale factor is preserved, the false alarm rate remains directly controllable as in conventional OS-CFAR. Compared with conventional OS-CFAR and U-Net-based algorithms, the proposed method maintains comparable detection performance while enhancing computational efficiency, suggesting its potential for CPU-based embedded implementations.