RailFDNet: A hybrid supervision and feature discrepancy enhancement model for railway anomalous object detection

Abstract

Detecting intrusive objects in track areas is essential for autonomous train operation. The diverse nature of intrusive objects and low occurrence probability make it hard to rely on object detection for this problem. In the face of this challenge, anomaly detection offers an effective solution. This work presents a novel model called RailFDNet for railway anomalous object detection, which has two main components. The first component is a backbone constructed using normalizing flow, optimized through unsupervised training via maximum likelihood estimation. The flow-based backbone can enhance the distinction between normal and abnormal data by transforming feature distribution. The second component is an anomaly decoder built on transformer, optimized through supervised training utilizing artificially generated anomalies. Its objective is to achieve more accurate anomaly localization at the local scale by further clarifying the differences between normal and abnormal features. The final detection results are achieved by integrating the normalized features from the backbone with the decoder output. This fusion strategy can significantly enhance anomaly detection accuracy. Extensive experiments conducted on the railway anomaly detection dataset demonstrate the effectiveness of our method. RailFDNet achieves 97.46% image-level AUROC and 98.95% pixel-level AUROC, surpassing the performance of existing state-of-the-art method. Moreover, the experimental results on the MVTec AD dataset demonstrate that RailFDNet has good generalization, while intrusion experiments conducted under simulated rainfall conditions in railway scenarios show strong robustness of our method.