Abstract: This paper proposed a target detection method based on improved YOLO (You OnlyLook Once) v8 to address the difficulty of detecting intrusive objects in metro low-illumination environments. This method significantly improved the feature extraction ability of low contrast targets by integrating RetinexFormer low light enhancement network and LSKA (Large Scale Kernel Attention) module, while maintaining lightweight and enhancing detection performance in complex scenes. The paper conducted experimental verification on a self-constructed subway low light intrusion object dataset, and the results showed that the improved Retinexformer LSKA-YOLOv8n model achieved 0.839 on the mAP50-95 index, which was about 9.24% higher than the original YOLOv8n model and 32.19% higher than the traditional Faster R-CNN model. This model has significantly improved recognition performance and can accurately detect intrusion objects in low light scenes of the subway, provide technical support for the safe operation of the metro.