Dyna-OpenVINS: Dynamic Environment Visual-Inertial Odometry with Object Removal via Multi-Scale Semantic Segmentation

In many Simultaneous Localization and Mapping (SLAM) systems, the assumption of a static environment is a common prior, which limits their performance in highly dynamic scenes. This paper introduces Dyna-OpenVINS, a lightweight Visual-Inertial odometry system built on the OpenVINS framework, specifically tailored for dynamic environments. Dyna-OpenVINS features a dynamic object removal module based on multi-scale semantic segmentation using the YOLOv8-seg model, effectively filtering out feature points associated with dynamic objects and mitigating their adverse effects. Additionally, the system is highly adaptable, supporting any number of cameras and incorporating a cross-camera feature tracking method to minimize redundant landmarks. We evaluated Dyna-OpenVINS using the dynamic VIODE dataset and in real-world scenarios. The results show that Dyna-OpenVINS achieves high localization accuracy while maintaining real-time performance, making it highly effective in dynamic environments.