Research on Visual SLAM Algorithms in Dynamic Environments

In dynamic environments, dynamic objects pose significant challenges to visual Simultaneous Localization and Mapping (SLAM) algorithms. They can lead to positioning drift, degradation of the map, and in severe cases, system failure. Moreover, the point features extracted by the Features from Accelerated Segment Test (FAST) suffer from inadequate geometric constraints in certain special environments. This deficiency further undermines the robustness of the algorithm. To tackle these issues, this research combines line features with the YOLO algorithm based on ORB-SLAM3 to form a SLAM algorithm (OYL-SLAM), which integrates point and line features with the YOLOv8 framework. Initially, the Line Segment Detector (LSD) algorithm is employed to extract line segment features. Given the prevalence of structured lines in artificial scenes, these line features are used to offset the limitations of point features. Subsequently, the YOLOv8s-seg semantic segmentation network is leveraged to identify dynamic objects within the environment. Once detected, the point and line features associated with these dynamic objects are removed. This study capitalizes on the complementary nature of point and line features, along with the dynamic feature detection and elimination capabilities of YOLOv8, to enhance the robustness of the SLAM system in dynamic environments. Simulation experiments using the TUM RGB - D public dataset demonstrate the efficacy of the OYL SLAM algorithm. Compared to the original algorithm, it reduces the average absolute pose error by 77.46% and the average relative pose error by 40.23%. Besides ensuring high-precision positioning, the OYL-SLAM algorithm is capable of generating a more detailed environmental map, thereby showcasing enhanced robustness.