Non-specific obstacle detection in the railway track area based on deep learning

Abstract Train operation safety is vital for railway systems, with obstacles intruding into railway tracks posing significant risks. However, existing research often focuses on specific obstacles, limiting the ability to detect a wide range of potential hazards like landslides or uncommon animals. Furthermore, many methods do not define the intrusion area, leading to excessive computational burden. Therefore, this paper detects non-specific obstacle (excluding humans and trains) that may affect the train operation safety and utilizes a segmentation network to delineate the area of obstacle intrusion. We propose two methods based on deep learning, namely a stepwise detection method (SDM) and an integrated detection method (IDM). The SDM uses a segmentation network to define track areas and reduce detection frequency, achieving 94% [email protected] for detection and 87% mIoU for track segmentation. In contrast, the IDM employs parallel multi-task networks for more accurate detection in complex areas, achieving 94% [email protected] and 89% mIoU, but at a higher computational cost. Through a comparative analysis, this research shows that the SDM is suitable for simpler track areas with lower computational requirements, while the IDM is optimal for complex and crucial track areas requiring high detection accuracy and real-time performance. This study provides new strategies for the application of railway obstacle detection method to the actual railway environment.