FM-SLAM: A Motion-Aware Visual SLAM Approach for Dynamic Environments Using Fast-SAM and Geometric Consistency

Abstract Accurate visual SLAM in dynamic environments remains a critical challenge due to feature misassociations caused by moving objects. Traditional methods like ORB-SLAM2 often fail to differentiate dynamic and static features, leading to degraded pose estimation and mapping integrity. This paper introduces FM-SLAM, a novel framework that syn-ergizes the zero-shot instance segmentation capability of Fast-SAM with motion-guided geometric consistency checks to eliminate dynamic interference. Unlike prior approaches reliant on semantic priors or manual annotations, FM-SLAM first identifies temporally consistent motion regions using geometric epipolar constraints and refines these regions via Fast-SAM to generate precise dynamic object masks. Feature points within these masked areas are discarded during pose optimization, ensuring robust static feature utilization. Key innovations include: (1) a lightweight, training-free mo-tion-consistency validation strategy that eliminates dependency on prior semantic knowledge, and (2) seamless integration of geometric cues and instance segmentation for enhanced generalizability in unknown environments. Extensive evalua-tions on the TUM RGB-D dataset demonstrate FM-SLAM’s superiority: it achieves 95.76% reduction in absolute trajectory error (ATE) and 93.62% lower relative pose error (RPE) compared to ORB-SLAM2 in high-dynamic sequences. The framework operates in real time at 25 FPS on a single GPU, showcasing practical viability for robotics and augmented real-ity applications. Project page: https://github.com/qiaoyang-adxs/FM-SLAM.git.