Separated-SLAM: a visual-inertial SLAM with adaptive atlas fusion and tracking mode switching for indoor scenes

We present Separated‑SLAM, a visual–inertial framework for effective indoor scene recognition and tracking. A new Atlas Evaluator thread evaluates the sparse indoor map for scene recognition and scene matching. For scene recognition, we preprocess sparse map points, project them onto three planes with a topology‑preserving intensity model, and encode the projections as compact histograms for fast matching. The matching results allow Separated‑SLAM to load the corresponding atlas from the database in real time while SLAM is running. After loop‑closure detection, the atlases are merged and the system switches to only‑tracking. During only‑tracking, ORB parameters are dynamically adjusted via the mode decision module to reduce the cost of feature extraction while preserving accuracy. Experiments on TUM‑VI show that Separated‑SLAM achieves accuracy comparable to ORB‑SLAM3 while reducing computation runtime by about 30% on average over complete trajectories; fine‑grained profiling further shows a 56.9% per‑frame reduction in the stabilized only‑tracking regime relative to full SLAM. The approach substantially lowers computational overhead when an atlas is available, streamlines sparse‑map matching, and provides a practical strategy to accelerate tracking after atlas merging.