Dynamic Visual SLAM Algorithm Based on Lightweight YOLOv8

Visual SLAM technology is one of the key issues in autonomous driving and mobile robots. Traditional algorithms often reduce the robustness of the system when facing highly dynamic environments and lack visual semantic information. The main purpose of this paper is to solve the problem of eliminating dynamic feature points based on semantic information in highly dynamic environments, so that the system can build maps for static feature points. This paper mainly proposes a real-time detection method of dynamic monitoring and semantic segmentation of visual information based on YOLOv8 to achieve dynamic feature point elimination in dynamic environments. First, YOLOv8 is applied to perform synchronous dynamic detection and semantic segmentation on the input image, and the target box of the target detection result is used to eliminate dynamic feature points. The SLAM system is then used to estimate the posture of static feature points. Finally, the point cloud library is used to construct a semantic map, thereby generating a semantic map that can be applied to actual scenes.