Visual-based Simultaneous Localization and Mapping has been extensively studied and applied to navigation in unmanned aerial vehicles. However, SLAM remains challenging in dynamic environments, where moving objects may be misidentified as static landmarks, compromising pose estimation. While many data-driven solutions have been proposed to address this issue, they often come at the cost of real-time performance, particularly on resource-constrained platforms. This paper presents a geometry-based approach to prune non-stationery features with minimal computational overhead. Features corresponding to the same spatial point are matched and tracked across consecutive frames, with epipolar errors computed over multiple frame pairs. Features exhibiting epipolar consistency over time are classified as stationary features and are constructed into map points. The proposed method is implemented into the state-of-the-art ORB-SLAM3 framework for evaluation. Dataset experiments demonstrate that the proposed SLAM system outperforms the baseline in both mapping efficacy and localization accuracy, while maintaining real-time performance without GPU acceleration. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
原文链接
分享
引用
