To address the low accuracy of existing nighttime airport runway foreign object detection algorithms, this paper proposes the Low-Light YOLOv5s FOD algorithm by integrating low-light enhancement and YOLOv5 detection concepts. During the training of the low-light enhancement network, the brightness layer channel, normalized for brightness, is incorporated into the attention mechanism, allowing the model to focus more on dark areas. Subsequently, dark-light FOD images undergo low-light enhancement. On the YOLOv5s detection network, multiscale feature fusion, a small object detection layer, and the NWD loss function are employed to enhance small object detection and address position sensitivity issues. ODConv, a full-dimensional dynamic convolution, replaces the standard convolution, further improving accuracy with multidimensional complementary attention mechanisms. Finally, foreign object detection is performed on enhanced images. Experimental results indicate a 12.6% improvement in NIQE scores for the improved EnlightenGAN restored images and a 4.3% increase in mAP for the improved YOLOv5s compared to the original model. In nighttime environments, the proposed algorithm achieves an average detection accuracy of 99.39%, a 67.99% improvement over the original algorithm without low-light enhancement, at a detection speed of 50.30 frames/s. Balancing accuracy and real-time performance, this algorithm effectively addresses false positives and misses in FOD detection during nighttime conditions.