Dynamic SLAM Using Video Object Segmentation: A Low Cost Setup for Mobile Robots

SLAM approaches, whether traditional, deep learning-based, or incorporating radiance field representations, face a common challenge in handling dynamic scenes due to their assumption of a static environment. This paper introduces a dynamic SLAM approach that integrates and enhances various SLAM methodologies. The primary focus lies in utilizing video object segmentation to bolster SLAM performance in dynamic environments, while concurrently developing an affordable mobile robot system tailored for such applications. While some methods have attempted dynamic SLAM, they often rely on single-object semantic segmentation, which is effective only for known object classes with prior knowledge of their static or dynamic nature. To address this limitation, we propose incorporating video object segmentation methods into our approach, which combines segmentation and tracking. Additionally, we construct a low-cost mobile robot system for dynamic SLAM applications and curate a dataset using affordable RGB-D sensors. To demonstrate the effectiveness of our method, we conduct an experimental study using the TUM-RGBD dataset and our RoverLab dynamic SLAM dataset, showcasing an enhancement in SLAM accuracy. Our system seamlessly integrates with both classical and modern Implicit SLAM algorithms, ranging from ORB-SLAM to Point-SLAM methods.