Source Detection in Broadband Passive SONAR with Vision Transformers and Poisson RFS Loss

Broadband passive SONAR systems must detect multiple acoustic sources in environments marked by high noise and limited prior information. Traditional model-based approaches, such as Cell Averaging Constant False Alarm Rate (CA-CFAR), rely on analytical models, which can fail to fully capture complex conditions, creating a need for data-driven methods. We propose a detection framework that integrates a Poisson Random Finite Set (RFS)-based loss function with a Vision Transformer (ViT) architecture. The ViT component processes Bearing Time History (BTH) waterfall data patches, capturing local and global acoustic features, while the Poisson RFS loss naturally accommodates an unknown number of sources. A test-time augmentation (TTA) strategy further boosts performance by exploiting the circular symmetry in bearing data. Experimental results show that our approach improves upon a conventional CFAR detector and CNN-based baselines across Receiver Operating Characteristic (ROC), Precision-Recall, and detection probability metrics. In particular, pairing ViT with the RFS loss yields higher accuracy and robustness to noise, all within a computationally feasible framework for real-time detection tasks.