Grid-based Fast and Structural Visual Odometry

TL;DR

GFS-VO is proposed, a grid-based RGB-D visual odometry algorithm that maximizes the utilization of both point and line features and incorporates fast line extraction and a stable line homogenization scheme to improve feature processing.

Abstract

In the field of Simultaneous Localization and Mapping (SLAM), researchers have always pursued better performance in terms of accuracy and time cost. Traditional algorithms typically rely on fundamental geometric elements in images to establish connections between frames. However, these elements suffer from disadvantages such as uneven distribution and slow extraction. In addition, geometry elements like lines have not been fully utilized in the process of pose estimation. To address these challenges, we propose GFS-VO, a grid-based RGB-D visual odometry algorithm that maximizes the utilization of both point and line features. Our algorithm incorporates fast line extraction and a stable line homogenization scheme to improve feature processing. To fully leverage hidden elements in the scene, we introduce Manhattan Axes (MA) to provide constraints between local map and current frame. Additionally, we have designed an algorithm based on breadth-first search for extracting plane normal vectors. To evaluate the performance of GFS-VO, we conducted extensive experiments. The results demonstrate that our proposed algorithm exhibits significant improvements in both time cost and accuracy compared to existing approaches.

Figures (7)

• Figure 1: Example of typical structural scene. Lines in the scene have parallel of perpendicular relationship with the axis of MA, which can be used for optimization.
• Figure 2: Overview of GFS-VO
• Figure 3: The comparison of plane normal extraction algorithm. Left is the result of our BFS-based algorithm while right is integral graph based algorithm. Compared with traditional method, our method can break the limit of regular grid and less affected by noise.
• Figure 4: The example of dense part. We use four color to represent nodes in quadtree during division. In the orange grid, algorithm choose the most representative line $A$ and filter line $C$. But in green grid that $A$ doesn’t passed, the most representative line become $C$. So algorithm choose line $C$ to retain and cause incomplete homogenization in orange grid.
• Figure 5: The illustration of search score expansion. Instead of searching in the surrounding grid, grids passed by line extension will also be searched.