SamSLAM: A Visual SLAM Based on Segment Anything Model for Dynamic Environment

With the thriving advancements of autonomous robotic agents, Simultaneous Localization and Mapping (SLAM) has increasingly gained importance in enhancing the mobility of intelligent robots over the years, owing to its capability to provide precise positional information and comprehensive environmental perception. However, the strong assumption of a static environment, which conventional SLAM approaches heavily rely on, imposes significant constraints on their practical applicability in real-world scenarios. Real-world scenarios are characterized by the presence of diverse and unpredictable dynamic targets. In this paper, to deal with these problems, a robust visual SLAM based on Segment Anything Model (SAM) named SamSLAM is proposed for dynamic environments. In SamSLAM, we leverage not only the accurate object boundary segmentation capability of SAM but also the geometric information inherent in the sequences of image data. The dynamic characteristics of different parts of the segmentation mask can be detected using the geometric information between consecutive images, which can be distinguished using techniques such as random sample consensus (RANSAC) and epipolar constraints. By combining the geometric information with SAM's segmentation mask, SamSLAM effectively reduces the impact of dynamic objects. We conduct experiments on two famous RGB-D dataset in SLAM field and in real-world environment. The results demonstrate that SAM-SLAM exhibits improved performance in dynamic environments. In addition, a parallel dense, colorful point cloud mapping thread is added, which provides the robot with richer environmental information.