| SLAM(Simultaneous Localization and Mapping)is a crucial problem in fields such as robotics and autonomous driving,with the objective of generating environmental maps and simultaneously determining the robot’s position within the map using sensor data.However,the performance of SLAM methods is significantly limited by dynamic objects,as traditional SLAM approaches are based on the assumption of static environments.In this context,the following work has been conducted under the conventional SLAM framework.Firstly,this study addresses the issue of dynamic objects in the scene by introducing a semantic mask,effectively reducing the impact of these objects on SLAM methods.Specifically,Deep Lab v3 is employed as the semantic segmentation network,with the PSA attention mechanism incorporated to enhance network performance.However,semantic segmentation reduces the number of acquired feature points and results in nonuniform distribution of feature points in the images.To tackle this issue,line segment features are introduced as a supplement.An end-to-end line segment extraction network is designed,which is better adapted to the environment and yields more robust feature lines compared to traditional extraction algorithms.This approach addresses the problems of fragmented and difficult-to-match line segment features in conventional algorithms.Finally,the proposed algorithms are applied to SLAM methods and experimentally validated on public datasets.The experimental results indicate that the designed method can effectively mitigate the impact of dynamic objects on SLAM methods.Consequently,the research findings hold potential application value for addressing SLAM problems in dynamic environments.In summary,this study reduces the influence of dynamic objects on SLAM methods and improves the robustness and accuracy of SLAM algorithms by introducing semantic masking techniques and an end-to-end line segment extraction network.This research provides a novel approach and methodology for solving SLAM problems in dynamic environments. |
