RailEINet:A novel scene segmentation network for automatic train operation based on feature alignment

Abstract

The primary prerequisite for realizing automatic train operation is endowing trains with the capability of independent environmental perception. The railway scene is notably intricate, encompassing elements such as tracks, poles and more. Scene segmentation aims to make a pixel-wise classification for a full perspective analysis of railway scene, which is geared to build a powerful automatic train perception system. The existing methods primarily emphasize the creation of multi-scale feature interaction mechanisms, where features at different levels are aggregated after the sampling operation. This operation neglects the differences in data among various features, which leads to the production of semantically ambiguous and unaligned features. This can significantly impact the segmentation results. To tackle this problem, we design two neural modules. Concretely, the Explicit Boundary Alignment (EBA) module is designed to utilize edge supervision to constrain direct alignment within the boundary regions among objects. This enables the refinement of edge details. Then, the Implicit Pyramid Alignment (IPA) module is designed to dynamically learn an offset map. This map, when combined with bilinear sampling operations, effectively mitigates the misalignment issues between multi-scale features. The two modules described above constitute a novel scene segmentation network tailored for railway scene perception, known as the Rail Scene-oriented Explicit-Implicit Feature Alignment Network (RailEINet). Extensive experiments are conducted to demonstrate the effectiveness of RailEINet. In particular, we achieve 66.22% mIoU on the wildly-used RailSem19 dataset and experimental results show that RailEINet can achieve excellent segmentation of various targets in railway scenarios.