SSCD-YOLO: Semi-Supervised Cross-Domain YOLOv8 for Pedestrian Detection in Low-Light Conditions

To mitigate the domain differences between infrared and visible light images and address the challenges of data annotation and poor pedestrian detection performance in low-light environments, we propose a semi-supervised cross-domain YOLOv8 pedestrian detection model SSCD-YOLO for low-light environments. First, a CDF-CycleGAN cross-domain image fusion method is designed, which improves the input of CycleGAN using a diffusion model, integrates the self-attention mechanism into the residual network of the CycleGAN generator to design the SA-Block, replaces the linear layer in the CycleGAN discriminator with an autoencoder, and improves the training method of CycleGAN. These enhancements achieve complementary information exchange between infrared and visible images, thereby improving the robustness and accuracy of the cross-domain object detection model. Second, the Mean Teacher approach is combined with the YOLOv8 detector, the EMA module is added to the output part of the YOLOv8 detection head, and the Swin Transformer block is used to enhance the C2f module of YOLOv8. Additionally, a cross-domain distillation loss function is designed to improve the performance and efficiency of multi-scale object detection. Experimental results show that CDF-CycleGAN outperforms CycleGAN on various metrics across the LLVIP and M<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>FD datasets. The overall detection performance of SSCD-YOLO on both LLVIP and M<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>FD datasets surpasses other comparative algorithms while meeting real-time detection requirements.