Intelligent foreign object detection on ballastless track beds via multimodal learning

The detection of foreign objects on ballastless track beds is critical for ensuring railway operational safety, yet existing manual inspection methods suffer from inefficiency and delayed hazard resolution. This paper proposes a novel multimodal fusion learning framework to address the challenges of detecting diverse and unpredictable foreign objects under high-speed image acquisition conditions. Leveraging deep learning advancements, our method employs contrastive learning to align normal track bed images with text features in the latent space, enabling anomaly detection through feature deviation analysis in abnormal scenarios. Experimental results on a self-collected dataset demonstrate the method's competitive performance, offering a practical solution for intelligent railway inspection tasks. The approach eliminates reliance on exhaustive anomaly data by training solely on normal samples, effectively overcoming limitations in recognizing rare or undefined foreign object categories.

Duke Scholars

Author Xin Li Electrical and Computer Engineering