Answers are grounded in this paper and cite their sources.
Visual simultaneous localization and mapping (SLAM) has been widely applied in indoor robotic navigation, augmented reality, and intelligent systems, with potential applications in educational robotics, interactive teaching environments, and spatially aware learning systems. However, most traditional visual SLAM systems rely on the assumption of scene rigidity, which makes them vulnerable to dynamic objects in indoor environments. To address these problems, this paper proposes an RGB-D SLAM system for indoor dynamic environments based on adaptive triangulation-affine local consistency and semantic filtering. Specifically, first, an adaptive triangulation-affine local consistency matching strategy (ATALC-matching) is designed to detect outliers and optimize the matching set. Reliable local structures are evaluated using a barycentric triangulation constraint, while degenerate regions are further verified using an affine fallback model. Second, an efficient semantic segmentation model is employed to obtain object-level regions, and dynamic objects are identified according to the ratio of outliers within each segmented region. Finally, both the outliers detected by ATALC-matching and the features located in dynamic regions are removed from the matching set before visual pose estimation. Meanwhile, map points associated with detected outliers are erased during map updating to reduce the accumulation of unreliable landmarks in the global map. Extensive evaluations are conducted on the TUM RGB-D dataset, the Bonn RGB-D Dynamic Dataset, and real-world scenarios. Compared with ORB-SLAM3, the proposed system reduces the root-mean-square-error of the absolute trajectory error by 94% and 38% on the highly and slightly dynamic TUM sequences, respectively. Evaluation on runtime shows that the complete system achieves an average processing speed of approximately 20 FPS. Real-world evaluations further demonstrate the effectiveness of the proposed methods in detecting outliers and dynamic objects.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: