Physics-Informed Generative Adversarial Networks for Range-Doppler Map Generation under Inter-Vehicle Occlusion

Deep Learning has recently expanded the potential of radar technology. Integrating neural networks with radar enables automatic feature extraction from radar images, enhancing target detection accuracy. However, these models require substantial volumes of labeled radar data for effective feature discernment, and the labeling process is resource-intensive, time-consuming, and costly due to its reliance on manual labor or complex simulation. Generative Adversarial Network (GAN) has emerged as a promising alternative for generating synthetic radar data. Existing studies have primarily focused on using GAN to generate single-target radar data, ignoring occlusion inconsistencies caused by multiple targets, which affects the realism and usability of the synthetic Range-Doppler Maps (RDMs). As a solution to this problem, this work utilizes a Physics-Informed GAN (PIGAN) that integrates occlusion modeling into the GAN architecture and incorporates a physics-aware loss function based on occlusion degree consistency. The proposed method ensures that the synthetic RDMs more accurately reflect the detection characteristics of simulation data under varying occlusion levels.