测绘学报 ›› 2019, Vol. 48 ›› Issue (6): 708-717.doi: 10.11947/j.AGCS.2019.20180421

• 摄影测量学与遥感 • 上一篇    下一篇

特征法视觉SLAM逆深度滤波的三维重建

张一, 姜挺, 江刚武, 余岸竹, 于英   

  1. 信息工程大学地理空间信息学院, 河南 郑州 450001
  • 收稿日期:2018-09-06 修回日期:2019-02-26 出版日期:2019-06-20 发布日期:2019-07-09
  • 作者简介:张一(1989-),男,博士生,研究方向为数字摄影测量与视觉SLAM。E-mail:276690308@qq.com
  • 基金资助:
    国家自然科学基金(41501482;41471387;41801388)

3D reconstruction with inverse depth filter of feature-based visual SLAM

ZHANG Yi, JIANG Ting, JIANG Gangwu, YU Anzhu, YU Ying   

  1. Institute of Surveying and Mapping, Information Engineering University, Zhengzhou 450001, China
  • Received:2018-09-06 Revised:2019-02-26 Online:2019-06-20 Published:2019-07-09
  • Supported by:
    The National Natural Science Foundation of China (Nos. 41501482;41471387;41801388)

摘要: 针对现有特征法视觉SLAM只能重建稀疏点云、非关键帧对地图点深度估计无贡献等问题,本文提出一种特征法视觉SLAM逆深度滤波的三维重建方法,可利用视频序列影像实时、增量式地构建相对稠密的场景结构。具体来说,设计了一种基于运动模型的关键帧追踪流程,能够提供精确的相对位姿关系;采用一种基于概率分布的逆深度滤波器,地图点通过多帧信息累积、更新得到,而不再由两帧三角化直接获取;提出一种基于特征法与直接法的后端混合优化框架,以及基于平差约束的地图点筛选策略,可以准确、高效解算相机位姿与场景结构。试验结果表明,与现有方法相比,本文方法具有更高的计算效率和位姿估计精度,而且能够重建出全局一致的较稠密点云地图。

关键词: 视觉即时定位与地图构建, 三维重建, 逆深度滤波器, 运动模型, 后端混合优化框架

Abstract: Aiming at the problem that the current feature-based visual SLAM can only reconstruct a sparse point cloud and the ordinary frame does not contribute to point depth estimation, a novel 3D reconstruction method with inverse depth filter of feature-based visual SLAM is proposed, which utilizes video sequence to incrementally build a denser scene structure in real-time. Specifically, a motion model based keyframe tracking approach is designed to provide accurate relative pose relationship. The map point is no longer calculated directly by two-frame-triangulation, instead it is accumulated and updated by information of several frames with an inverse depth filter based on probability distribution. A back-end hybrid optimization framework composed of feature and direct method is introduced, as well as an adjustment constraint based point screening strategy, which can precisely and efficiently solve camera pose and structure. The experimental results demonstrate the superiority of proposed method on computational speed and pose estimation accuracy compared with existing methods. Meanwhile, it is shown that our method can reconstruct a denser globally consistent point cloud map.

Key words: visual simultaneous localization and mapping, 3D reconstruction, inverse depth filter, motion model, back-end hybrid optimization framework

中图分类号: