关键词: 铁路工程； 轨道异物； 轻量化； 双向特征融合； 注意力机制； 聚集损失

Abstract: To address the issues of low target recognition accuracy and poor detection real-time performance in the task of railway foreign object intrusion limit detection, an orbital foreign object intrusion limit detection model combining bidirectional feature fusion with focusing loss was proposed. The lightweight S-GhostNet was used as the backbone network to extract features, aiming to reduce computational complexity and enhance processing speed. A bidirectional feature fusion structure (Sim-DFPN) was designed. The non-parametric attention mechanism was introduced to enable the model to focus more on the feature information of the target region, which effectively suppressed interference from invalid background noise and enhanced the flow of information between different levels. Furthermore, in order to further improve detection performance, the intersection over union (IoU) loss was restructured. A linearly spaced method was used to design an adaptive focusing edge loss function, Focaler EIoU, and experiments were conducted on the railway foreign object intrusion limit dataset. The research results show that the proposed network achieves a detection accuracy of 90% on this dataset, with a model size of only 55 MB and a frame rate of 79 frames per second.

Key words: railway engineering; orbital foreign objects; lightweight; bidirectional