Improving Robustness in Dynamic Scenes: A Framework for Visual-Inertial-Semantic SLAM with Dynamic Probability Fusion

With the ongoing advancement of intelligent robotics, vision-based SLAM (Simultaneous Localization and Mapping) technology plays a pivotal role in various fields. Traditional SLAM systems, often predicated on static environment assumptions, demonstrate limited robustness and accuracy in highly dynamic settings. In response, this study proposes a dynamic probability update algorithm aimed at improving the camera tracking accuracy and robustness in dynamic environments. By Integrating the BiSeNet V2 semantic segmentation network with the IMU pre-integrated re-projection model, the proposed method effectively identifies and excludes dynamic objects, thus ensuring more accurate estimations of static environmental features. Furthermore, we propose a keyframe processing mechanism that eliminates dynamic point clouds in keyframes, facilitating the construction of stable static maps. It is also improved based on ORB-SLAM3 and experimented on the TUM RGB-D dataset. The results show that the proposed system improves the localisation accuracy and robustness of visual SLAM in highly dynamic environments.