SFSE-SLAM: Semantic-Geometric Filtering and Texture-Aware Compensation for Robust Visual SLAM in Dynamic Indoor Scenes

Visual simultaneous localization and mapping in indoor dynamic environments remains challenging because moving objects introduce unreliable correspondences, whilst removing dynamic feature points often leaves insufficient static features in low-texture regions. This paper proposes SFSE-SLAM (Semantic Feature Selection and Enhancement SLAM), a robust visual SLAM framework that combines semantic-geometric feature filtering with texture-aware feature compensation to improve pose estimation under dynamic interference. The framework first identifies potentially dynamic regions through pixel-level semantic segmentation and removes features associated with highly dynamic objects. To reduce over-filtering and address semi-static objects, depth variation and multi-view geometric consistency are further used to distinguish static and moving feature points across consecutive frames. After dynamic filtering, a learned local feature extractor is introduced to improve descriptor discriminability and feature density in reliable static regions. Two additional modules, semantic confidence weighting and static region feature compensation, adaptively adjust feature extraction thresholds so that low-texture but geometrically useful areas can contribute more stable correspondences. The proposed system is evaluated on public dynamic RGB-D benchmarks, including the TUM RGB-D dataset (featuring high-dynamic human motions) and the Bonn dataset (containing complex non-human dynamic objects such as thrown balloons and moved boxes). Experimental results indicate that the method improves localization robustness in high-dynamic scenarios and reduces trajectory error compared with conventional ORB-based SLAM and several dynamic SLAM baselines. Specifically, on the highly dynamic TUM fr3-walking-xyz sequence, our method achieves an ATE RMSE of 0.0148 m, representing a 97.9% reduction compared to ORB-SLAM3. The study demonstrates the potential of combining semantic priors, geometric verification and adaptive feature enhancement for dynamic indoor localization.