The working environment in underground coal mines is extremely complex, and a lightweight network fusion-based method for detecting foreign objects in underground tracks is proposed to address the problem of missed target detection and misdetection due to noise, dust, and shaded areas. First, targeted improvements are made to the YOLOv7-Tiny network framework by introducing the FasterNet and GhostV2 modules, which reduces the model complexity while maintaining the high performance level. Second, the ELAN_E module is designed in combination with the ECA (Efficient Channel Attention) attention mechanism, which improves the sensitivity of the model to foreign object features. Finally, the Focal-EIoU loss function is used to optimize the coordinate loss calculation, which further improves the detection accuracy. The experimental results show that compared with YOLOv7-Tiny, the proposed method reduces 30.23% and 15.15% in terms of the number of parameters and computation, respectively. Meanwhile, the mAP (mean Average Precision) index is improved by 1.2%, which effectively improves the accuracy and efficiency of foreign object detection, and provides an effective improvement scheme for the intrusion detection in the track area of underground coal mine.