DFT-VSLAM: A Dynamic Optical Flow Tracking VSLAM Method

Abstract

Visual Simultaneous Localization and Mapping (VSLAM) technology can provide reliable visual localization and mapping capabilities for critical tasks. Existing VSLAM can extract accurate feature points in static environments for matching and pose estimation, and then build environmental map. However, in dynamic environments, the feature points extracted by the VSLAM system will become inaccurate points as the object moves, which not only leads to tracking failure but also seriously affects the accuracy of the environmental map. To alleviate these challenges, we propose a dynamic target-aware optical flow tracking method based on YOLOv8. Firstly, we use YOLOv8 to identify moving targets in the environment, and propose a method to eliminate dynamic points in the dynamic contour region. Secondly, we use the optical flow mask method to identify dynamic feature points outside the target detection object frame. Thirdly, we comprehensively eliminate the dynamic feature points. Finally, we combine the geometric and semantic information of static map points to construct the semantic map of the environment. We used ATE (Absolute Trajectory Error) and RPE (Relative Pose Error) as evaluation metrics and compared the original method with our method on the TUM dataset. The accuracy of our method is significantly improved, especially 96.92% on walking_xyz dataset. The experimental results show that our proposed method can significantly improve the overall performance of VSLAM systems under high dynamic environments.