Deep Learning-Based Railway Foreign Object Intrusion Intelligent Perception Using Attention-Aggregated Semantic Segmentation

Foreign object intrusion detection (FOID) is one of the critical tasks to ensure the safe and efficient operation of trains. Semantic segmentation, which involves pixel-level recognition of images, has been widely studied in automatic driving obstacle avoidance. However, unlike road transportation, the operation speed of trains requires higher detection efficiency. The availability of mature railway scenario datasets is limited compared to road transportation datasets. Therefore, considering the complexity of operating scenarios with diverse and unpredictable foreign objects, this article proposes a boundary-assisted dual-branch attention semantic segmentation network (BDANet). BDANet completes accurate segmentation while reducing parameters, enabling real-time semantic recognition of the railway environment. A COCO-Stuff-Rail dataset extracted based on COCO-Stuff is constructed to guide model training. Then, an adaptive correction algorithm is introduced to fine-tune the BDANet, making it generalizable to diverse realistic environments. Ultimately, this article achieves end-to-end track extraction, open-set foreign object detection, and common foreign object identification using a unified process. To evaluate the superiority of BDANet, comparison, and ablation experiments are conducted on the COCO-Stuff-Rail. Visual segmentation and open-set detection results of a real-world scenario validate that the proposed process can bridge the gap between the training set and practical applications.