Dynamic Environments and Robust SLAM: Optimizing Sensor Fusion and Semantics for Wheeled Robots

Abstract

This paper proposes an approach to enhance the robustness and accuracy of visual simultaneous localization and mapping (SLAM) for ground wheeled mobile robots in dynamic environments. The proposed method incorporates encoder measurements to establish optimization constraints in bundle adjustment. To further improve robustness, a geometric technique utilizing KMeans clustering with epipolar constraints and the SegNet for semantic segmentation is employed to filter out features detected on moving objects. These modifications are integrated into the state-of-the-art SLAM system ORB-SLAM3 and demonstrate superior accuracy and real-time performance compared to the baseline approach. The effectiveness of the proposed method is demonstrated through multiple OpenLoris and IROS Lifelong SLAM competition scenarios.