Boundary Dynamic Perception Detection Network for Low-Light Railway Environment

With the increase in railway operation mileage, railway perimeter security is facing significant challenges. These challenges are especially pronounced in low-light conditions such as dusk, nighttime, or tunnels, where edge details are often lost due to insufficient light. Conventional detection networks rely heavily on precise boundary guidance of targets, resulting in a high rate of false alarms and omissions during on-site detection. To this end, this study proposes a target boundary dynamic perception detection network (BDP-Net) under railway low-light environments. First, by constructing a spatial feature enhancement module, the low-light image is enhanced in the whole domain to strengthen the feature difference between the target and the background. Second, a parallel spatial enhancement convolutional two-branch network structure is used to retain the spatial information and generate the high-resolution features, and the two-branch network features are fused across the layers. Finally, a boundary dynamic representation module is proposed to process the fused feature information to enhance the information richness of the detected targets and improve the flexibility of regression. Extensive experiments on the RailDark and ExDark datasets demonstrate that the proposed BDP-Net achieves a 6.9% accuracy improvement on RailDark (reaching 80.1%) and a 2.1% accuracy improvement on ExDark (reaching 83.3%) compared to other detection networks (e.g., PEYOLO). The proposed algorithm can improve the accuracy and real-time performance of railway monitoring in low-light conditions, helping to prevent safety hazards and reduce accidents.