RLD-SLAM: A Robust Lightweight VI-SLAM for Dynamic Environments Leveraging Semantics and Motion Information

Existing mainstream dynamic simultaneous localization and mapping (SLAM) can be categorized into image segmentation-based and object detection-based methods. The former achieves high accuracy but suffers a heavy computational burden, while the latter operates at higher speeds but with lower accuracy. In this article, we propose robust lightweight dynamic SLAM (RLD-SLAM), a robust lightweight visual-inertial SLAM for dynamic environments, leveraging semantics, and motion information. Our novel approach combines object detection and Bayesian filtering to maintain high accuracy while quickly acquiring static feature points. In addition, to address the challenge of semantic-based dynamic SLAM in highly dynamic scenes, RLD-SLAM leverages motion information from the inertial measurement unit to assist in tracking dynamic objects and maximizes the utilization of static feature in the environment. We conduct experiments applying our proposed method on indoor, outdoor datasets, and unmanned ground vehicles. The experimental results demonstrate that our method surpasses the current state-of-the-art algorithms, particularly in highly dynamic environments.