作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (6): 201-207. doi: 10.19678/j.issn.1000-3428.0064598

• 图形图像处理 • 上一篇    下一篇

基于点面特征融合的RGB-D SLAM算法

孙超, 朱勇杰, 余林波, 苗隆鑫, 曹勉, 叶力, 郭乃宇   

  1. 江汉大学 智能制造学院, 武汉 430056
  • 收稿日期:2022-05-02 修回日期:2022-07-24 发布日期:2023-06-10
  • 作者简介:孙超(1984-),男,副教授、博士,主研方向为图像处理、数据处理;朱勇杰、余林波(通信作者)、苗隆鑫、曹勉、叶力、郭乃宇,硕士研究生。
  • 基金资助:
    国家重点研发计划(2018YFD1100104);湖北省重点研发计划(2020BCA084);湖北省重点培育学科控制科学与工程基金(2020XK015);江汉大学校级科研基金(2021yb153)。

RGB-D SLAM Algorithm Based on Point-Surface Feature Fusion

SUN Chao, ZHU Yongjie, YU Linbo, MIAO Longxin, CAO Mian, YE Li, GUO Naiyu   

  1. School of Intelligent Manufacturing, Jianghan University, Wuhan 430056, China
  • Received:2022-05-02 Revised:2022-07-24 Published:2023-06-10

摘要: 传统的视觉SLAM算法主要依赖于点特征来构建视觉里程计,而在人造环境中通常存在一些点特征不足但结构特征丰富的场景,此时基于点特征的视觉里程计方法容易出现跟踪失败或位姿估计精度降低的问题。提出一种融合点面特征的RGB-D SLAM算法,利用人造环境中的平面特征提高SLAM算法定位与建图的精度与鲁棒性。使用AGAST特征点提取算法并采用四叉树的方式进行改进,使特征点在图像中均匀分布以减少点特征提取的冗余度。同时,在传统点特征方法的基础上添加平面特征,使用连通域分割算法从点云中获取平面特征,并构建伪平面特征,结合AGAST特征点构建点面特征融合的结构约束因子图,添加多重约束关系用于图优化。实验结果表明,该算法AGAST特征点提取效率优于ORB-SLAM2算法,融合的点面特征使其在室内环境下的定位和建图精度更高,绝对轨迹误差减小约20%,相对轨迹误差减小约10%,单帧跟踪耗时减少约7.3%。

关键词: RGB-D SLAM算法, AGAST算法, 点云分割, 点面特征融合, 图优化

Abstract: Traditional visual SLAM algorithms rely mainly on point features to construct visual odometry. There are usually scenes with insufficient point features but rich structural features in artificial environments.Currently,visual odometry methods based on point features are prone to tracking failures or reduced pose estimation accuracy.Therefore, considering the above problems,an RGB-D SLAM algorithm that fuses point and surface features is proposed,which makes full use of the plane features in an artificial environment to improve the accuracy and robustness of the SLAM algorithm positioning and mapping.The AGAST feature point extraction algorithm is used and the quadtree is used to improve it,so that the feature points are evenly distributed in the image to reduce the redundancy of point feature extraction.On the basis of the traditional point feature method,the plane feature is added,and the plane feature is obtained from the point cloud using the connected domain segmentation algorithm,and the pseudo-plane feature is constructed.Combined with the AGAST feature points,a structural constraint factor graph of point-surface feature fusion is constructed,and multiple constraint relationships are added for graph optimization. Compared with the ORB-SLAM2 algorithm,the AGAST feature point extraction efficiency of the proposed algorithm is better than that of the ORB-SLAM2 feature extraction strategy,and the fusion of point and surface features makes the positioning and mapping accuracy better in indoor environments.The dataset test results show that the absolute trajectory error of the proposed algorithm is reduced by approximately 20%,the Relative Trajectory Error(RTE) is reduced by approximately 10%, and single-frame tracking time is reduced by approximately 7.3%.

Key words: RGB-D SLAM algorithm, AGAST algorithm, point cloud segmentation, point-surface feature fusion, graph optimization

中图分类号: