Dynamic point line SLAM method based on semantic segmentation

In indoor dynamic environments, especially in low-texture environments, traditional visual Simultaneous Localization and Mapping (SLAM) algorithms are subject to a large number of dynamic interferences and lack feature points for pose estimation, resulting in reduced positioning accuracy and poor robustness of the system. To tackle the aforementioned challenge, a novel dynamic point-line SLAM system that incorporates semantic segmentation, known as ODY-SLAM, has been introduced. This algorithm introduces an improved Line Segment Detector (LSD) algorithm on the basis of the ORB- SLAM3 framework and adds a dynamic segmentation and detection thread simultaneously. The dynamic segmentation and detection thread utilizes the YOLOv8 network to complete the semantic detection and segmentation in dynamic scenes. It divides the camera images into recognizable regions and unrecognizable regions, assigns dynamic semantic information to the recognizable regions and generates dynamic masks. Subsequently, the dynamic masks are cleansed of their characteristic points and linear elements. In addition, a geometric constraint technique is utilized to filter out feature points and lines in areas that are not identifiable. This algorithm is verified through experiments using the publicly available TUM dataset. The experimental outcomes indicate that the ODY-SLAM algorithm outperforms the ORB-SLAM3 algorithm in terms of pose positioning accuracy within indoor dynamic settings, showing an average enhancement of 68.55%, with the highest recorded improvement at 96.37%.