Hybrid Semantic SLAM Framework Based On Open Vocabulary

We introduce a hybrid open-vocabulary semantic SLAM framework that fuses precise geometric positioning with real-time semantic perception. This multi-threaded framework comprises an ORB-SLAM2 geometry thread, an asynchronous semantic thread consisting of Grounding DINO and SAM, and a fusion module that maps 2D semantics to 3D cuboid instances. Data test in the TUM RGB-D and EuRoc MAV claim that this framework achieves centimeter-level accuracy and strong robustness in dynamic scenes, demonstrating that open-vocabulary semantics significantly enhances SLAM performance in tasks such as path planning and human-robot interaction.