In this paper, we propose an improved ORB-SLAM3 method for low-light dynamic environments. Initially, the Retinexformer image enhancement algorithm is used to preprocess the input image to enhance the stability in low light environment. Subsequently, by lightweight YOLOv8 detection network is utilized for dynamic region identification. Furthermore, the LSD algorithm is integrated to incorporate line features, enhancing the system’s robustness. Finally, epipolar constraint is employed to eliminate dynamic feature points within dynamic regions, thereby effectively preserving static feature points. Experimental results demonstrate a 45.7% reduction in the parameter count of the lightweight YOLOv8 model, achieving model lightweighting. Compared to the original ORB-SLAM3 algorithm, the proposed method exhibits an 87.37% improvement in absolute trajectory accuracy and a 39.81% enhancement in relative pose accuracy within dynamic environments, thereby significantly improving the algorithm’s accuracy. Moreover, the application of the Retinexformer image enhancement algorithm facilitates superior feature extraction, enabling the algorithm to effectively address low-light environments.