Real-Time Railway Scene Perception Based on Fast Reconvolution Skip Fusion and Dual Attention Ladder Feature Enhancement

The safety of railway operations is closely related to the state of the train operating environment. Railway scene perception methods can sense the train operating environment and provide technical support for auxiliary driving. However, existing railway scene perception methods require large computations and they are hard to balance between speed and accuracy. There is, therefore, an urgent need to develop a fast and accurate railway scene perception algorithm. To address the need for fast and accurate perception of railway scenes, we propose a real-time railway scene perception network (RRSPNet). Firstly, we develop a lightweight dual-branch backbone and a fast reconvolution skip fusion for rapid extraction of image features. Secondly, we propose a double attention ladder feature enhancement to further improve high-level semantic features and enhance the overall perception accuracy of the network. Finally, we construct two different railway scene data sets: a subway tunnel data set named Metroscapes and a high-speed railway data set named Railscapes. We test the proposed algorithm on both data sets and the public data set rectified linear unit. The results demonstrate its superiority.