CHAMELEON-SLAM: Adaptive Feature Selection and Uncertainty-Aware Matching for Robust Monocular Visual SLAM

Visual SLAM systems employ fixed feature extractors regardless of scene content. This is suboptimal: high-texture outdoor environments favor fast, sparse features, while low-texture indoor scenes or motion-blurred frames demand denser or more robust alternatives. We present CHAMELEON-SLAM, a visual SLAM system that adapts its feature extraction strategy to scene characteristics and incorporates per-match uncertainty into pose optimization. Our approach introduces two complementary components: (1) a lightweight scene classifier that selects among XFeat, ALIKED, and SuperPoint based on texture density, edge structure, and blur metrics, adding under 1 ms of overhead, and (2) a multi-scale descriptor consistency measure that estimates per-keypoint uncertainty and down-weights unreliable matches in bundle adjustment without requiring additional training. We integrate both components into ORB-SLAM3's tracking and local mapping threads. Experiments on KITTI and EuRoC demonstrate that adaptive feature selection alone reduces absolute trajectory error by 41–65% over ORB-SLAM3 depending on the dataset, uncertainty weighting provides 34–59%, and the combined system achieves 45–70% lower ATE (57% on average) while maintaining real-time performance at 32 FPS. The system also improves by 11–25% over the strongest fixed baseline with learned matching (SuperPoint + LightGlue).