GFC-Net: Geometry and Feature-Driven Fast Segmentation for Railway Track Safety Limit Detection

In response to the challenges of the railway track safety limit segmentation task, such as the long and slender structure of the track being vulnerable to background interference, the difficulty in modeling the multitrack intersection topology deformation, and the severe redundancy of network computation, this article proposes a fast segmentation network for railway track safety limit based on geometric adaptation and feature interlacing complementarity geometry- and feature-driven fast segmentation network (GFC-Net). This network integrates geometrically adaptive dynamic offset convolution (GADC), sliding window attention encoder (SWA-Encoder), and the collaborative segmentation network GFC-Net with global–local feature interleaving complementary decoder (GLIC-decoder). First, the GADC module adaptively adjusts the sampling position and weight of the convolution kernel through dynamic offsets and modulation scalars to precisely capture the geometric deformation characteristics of the curved and crossed areas of the track and suppress the interference of background noise such as gravel and vegetation. In addition, the SWA-Encoder employs a step-size overlapping convolution block embedding and periodic window offset strategy to eliminate local window spatial isolation while maintaining linear complexity, enhancing the continuous modeling capability of hundreds of kilometers of band-shaped tracks. Moreover, the GLIC-Decoder, based on a minimalist architecture of fully connected layers and bilinear interpolation, unifies multiscale feature resolution and dynamically allocates weights. Experiments on the self-built datasets RailTrackSeg (1176 images) and RailScapes (840 high-speed scene images) show that the method in this article achieves an average intersection and union ratio (mIoU) of 88.91% and 91.15%, respectively, in scenarios of dense track intersections, curve deformations, and dynamic interference, which provides a reliable solution for high-precision real-time segmentation in complex railway scenarios.