Virtual Vectors for Vision-Based Simultaneous Localization and Mapping System

Abstract

Visual SLAM estimates the camera's pose by tracking feature points between consecutive frames. Static feature points are essential for accurately estimating the camera's pose, whereas moving feature points, due to their inherent motion, can lead to inaccuracies. To address this issue, this paper introduces the concept of virtual vectors. By analyzing the magnitude changes of virtual vectors between consecutive frames, our method detects and removes moving feature points, ensuring that only static feature points are utilized for pose estimation. Furthermore, feature points at varying distances have different impacts on pose estimation. Distant feature points are more suitable for estimating the rotation matrix, while closer feature points are better suited for estimating the translation matrix. By analyzing the directional changes of virtual vectors between consecutive frames, our method decouples the pose matrix, allowing for the selection of feature points at appropriate distances to estimate the rotation and translation matrices accurately. Experiments conducted on the KITTI and TUM datasets demonstrate the effectiveness of the virtual vector approach and show an improvement in pose estimation accuracy.