3OFRR-SLAM: Visual SLAM with 3D-Assisting Optical Flow and Refined-RANSAC

EasyChair Preprint 7032

Abstract

To perform navigation or AR/VR applications on mobile devices, SLAM is expected to be with low computational complexity. But using feature descriptors restricts the minimization and lightweight of a SLAM system. In this paper, we propose a lightweight monocular SLAM system called 3OFRR-SLAM, which is precise, fast, and achieves real-time performance on CPU and mobile phones. It integrates a 3D-assisting optical flow tracker, uses a local map to provide prior information for optical flow, and improves the Lucas-Kanade algorithm, which makes data association fast and reliable. To further eliminate outliers of data association, we propose a novel Refined-RANSAC, improving the accuracy of camera pose estimation without taking much extra time cost. To further eliminate outliers of data association, we propose a novel Refined-RANSAC, improving the accuracy of camera pose estimation without taking much extra time cost. We evaluate our system on TUM-RGBD dataset and real-world data. The results demonstrate that our system obtains an outstanding improvement in both speed and accuracy compared with current state-of-the-art methods ORB-SLAM2 and DSO.Moreover, we transplant our system to an android-based smartphone and show the application of this system for augmented reality (AR).

Keyphrases: Fast Tracking, SLAM, visual localization

@booklet{EasyChair:7032,
  author    = {Yujia Zhai and Fulin Tang and Yihong Wu},
  title     = {3OFRR-SLAM: Visual SLAM with 3D-Assisting Optical Flow and Refined-RANSAC},
  howpublished = {EasyChair Preprint 7032},
  year      = {EasyChair, 2021}}