USD-SLAM: A Universal Visual SLAM Based on Large Segmentation Model in Dynamic Environments

摘要

Visual Simultaneous Localization and Mapping (SLAM) has been widely adopted in autonomous driving and robotics. While most SLAM systems operate effectively in static or low-dynamic environments, achieving precise pose estimation in diverse unknown dynamic environments continues to pose a significant challenge. This letter introduces an advanced universal visual SLAM system (USD-SLAM) that combines a universal large segmentation model with a 3D spatial motion state constraint module to accurately handle any dynamic objects present in the environment. Our system first employs a large segmentation model guided by precise prompts to identify movable regions accurately. Based on the identified movable object regions, 3D spatial motion state constraints are exploited to remove the moving object regions. Finally, the moving object regions are excluded for subsequent tracking, localization, and mapping, ensuring stable and high-precision pose estimation. Experimental results demonstrate that our method can robustly operate in various dynamic and static environments without additional training, providing higher localization accuracy compared to other advanced dynamic SLAM systems.