Dynamic Object-Aware SLAM: YoloV8 Integration and Enhanced Multi-View Geometry for Reliable Mapping in Dynamic Environment

This paper introduces YoloV8-SLAM, a novel approach that addresses the limitations of current Visual Simultaneous Localization and Mapping (VSLAM) systems by incorporating dynamic object identification and enhanced multi-view geometry techniques. By integrating an adaptive segmentation method that combines geometric motion and prospective motion information, our approach enables the effective segmentation of dynamic objects. YoloV8, a cutting-edge object detection algorithm, and Enhanced Multi-View Geometry technique are employed to handle low, medium, and high dynamic scenarios, while a well-matched point selection algorithm extracts high-speed motion information. Extensive experiments demonstrate the superiority of YoloV8-SLAM over established and state-of-the-art SLAM systems, achieving significant reductions in Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). Compared to ORB-SLAM3, YoloV8-SLAM achieves up to 89.1% and 88.0% reductions in ATE and RPE, respectively. It also outperforms DynaSLAM, ORB-SLAM2, and DM-SLAM, with reductions exceeding 39.8% in ATE and 66.4% in RPE. The proposed YoloV8-SLAM method presents an accurate and efficient solution for SLAM, applicable across various scenarios.