DSK-SLAM: A Dynamic SLAM System Combining Semantic Information and a Novel Geometric Method Based on K-Means Clustering

In dynamic simultaneous localization and mapping (SLAM) systems, the traditional multiview geometry method faces two main challenges. The first issue to address is how to select an appropriate threshold in order to enhance the accuracy of determining the dynamic or static (D-S) property of matched points. Second, for the unmatched points between adjacent frames, it is difficult to determine their D-S property. In addressing the aforementioned issues, this article first obtains the dynamic mask by using semantic segmentation model and removes the dynamic feature points located within these masks. For the remained matched feature points, the article proposes an adaptive threshold-setting method for epipolar constraint by analyzing the distribution characteristics of the epipolar line distances. Furthermore, to handle unmatched points, the article proposes a novel method called the KCP method that combines K-means clustering and epipolar constraint. This method effectively calculates the dynamic probability and eliminates feature points with high dynamic probabilities. Finally, this article proposes a system named dynamic semantic K-means clustering (DSK)-SLAM based on the orb-slam2 system, which integrates the aforementioned methods into a new dynamic removal module. Experiments demonstrate that compared with ORB-SLAM2, DS-SLAM, DynaSLAM, and MOR-SLAM, DSK-SLAM acquire better results on the public dataset TUM RGB-D, which achieves an average accuracy improvement of 95.71%, 41.35%, 8.37%, and 1.75% in highly dynamic sequences, and achieves an accuracy improvements of 24.69%, 6.15%, 7.58%, and 15.28% in a low dynamic sequence, respectively.