We present an extension of ORB-SLAM3 that integrates object segmentation from a YOLOv8-seg model to enrich the map with semantic information. Segmentation is run in a separate process and the masks are passed to the SLAM system, where they are combined with depth data. Large background classes such as walls and floors are visualized as colored meshes, while static semantic objects like chairs and tables are also included in the optimization. For these objects, we add extra constraints that link their 3-D positions with camera poses and nearby points, so they become stable landmarks in the map. The result is a SLAM system that keeps the accuracy of the original ORBSLAM3 while providing a richer map with useful semantic objects. We also discuss current limits, such as occasional segmentation errors, and outline possible improvements with faster models and sensor fusion.
|
