Spatial-aware energy learning for out-of-distribution segmentation in railway intrusion detection.

Railway intrusion detection is crucial for ensuring the safety and reliability of railway operations. Deep learning-based railway intrusion detection (RID) methods struggle with unknown foreign objects, while existing pixel-level out-of-distribution (OoD) detection methods lack spatial awareness for railway scenes. To address this limitation, we propose an uncertainty-driven spatial-aware energy learning (SAEL) method for railway intrusion detection. Our approach introduces a spatial-aware energy learning strategy that incorporates a distance transform and a Gaussian smoothing weighting matrix to prioritize OoD pixels near the boundaries of in-distribution pixels. Additionally, we design a texture representation module to enhance the model's ability to capture edge and texture details of anomalous regions. We evaluate our method on both public datasets and a self-constructed dataset. Extensive experiments on both public datasets and our self-constructed OoD-RID test set demonstrate that our approach significantly outperforms state-of-the-art (SOTA) baseline methods. On the OoD-RID test set, the FPR95 is reduced by 10.68%, the AuPRC is improved by 5.40%, and the AuROC is increased by 1.73%. This work provides a robust and practical OoD segmentation solution for unseen intrusion detection in complex railway scenes, demonstrating a favorable balance between detection performance and computational efficiency for practical railway applications.