DynaLight-SLAM: A Visual SLAM of Dynamic Objects in Low-Light Conditions

After integrating the YOLO algorithm with ORB-SLAM3 while achieving good performance on dynamic datasets, its stringent requirements for lighting conditions have limited its applicability in unmanned aerial vehicles (UAVs). To address this limitation, this paper proposes a new visual SLAM system named DynaLight-SLAM, which significantly improves localization and tracking performance under low-light and dynamic conditions by integrating ORB-SLAM3 with a low-light correction module and YOLOv8. DynaLight-SLAM is built upon ORB-SLAM3, utilizing its powerful visual localization capabilities, and employs a low-light dynamic system that combines OpenCV and GAN networks to preprocess and enhance illumination in low-light images, thereby improving system robustness in dim environments. Additionally, YOLOv8 is used for semantic segmentation to generate masks that cover feature points within dynamic objects, reducing errors caused by such objects and extending the application range of ORB-SLAM3. Evaluation results on public datasets with added low-light noise and public low-light 3D modeling datasets demonstrate that DynaLight-SLAM maintains high computational efficiency while significantly enhancing robustness and reducing trajectory error. These features make DynaLight-SLAM an ideal choice for applications such as drone-based autonomous patrolling.