东北大学学报:自然科学版 ›› 2017, Vol. 38 ›› Issue (12): 1764-1768.DOI: 10.12068/j.issn.1005-3026.2017.12.020

• 机械工程 • 上一篇    下一篇

基于RGB-D的室内场景实时三维重建算法

胡正乙1,2, 谭庆昌1, 孙秋成3   

  1. (1. 吉林大学 机械科学与工程学院, 吉林 长春130022; 2. 长春汽车工业高等专科学校, 吉林 长春130013; 3. 长春师范大学, 吉林 长春130032)
  • 收稿日期:2016-07-18 修回日期:2016-07-18 出版日期:2017-12-15 发布日期:2018-01-02
  • 通讯作者: 胡正乙
  • 作者简介:胡正乙(1984-),男, 吉林长春人,吉林大学博士研究生,长春汽车工业高等专科学校讲师; 谭庆昌(1957-),男,吉林海龙人,吉林大学教授,博士生导师.
  • 基金资助:
    国家自然科学基金资助项目(51405184).

RGB-D Based Indoor Scene Real-Time 3D Reconstruction Algorithm

HU Zheng-yi1,2, TAN Qing-chang1, SUN Qiu-cheng3   

  1. 1. College of Mechanical Science & Engineering, Jilin University, Changchun 130022, China; 2. Changchun Automobile Industry Institute, Changchun 130013, China; 3. Changchun Normal University, Changchun 130032, China.
  • Received:2016-07-18 Revised:2016-07-18 Online:2017-12-15 Published:2018-01-02
  • Contact: HU Zheng-yi
  • About author:-
  • Supported by:
    -

摘要: 针对基于视觉的室内场景三维重建过程中存在三维点云匹配不准确、过程耗时和深度信息部分缺失的问题,提出一种带有深度约束和局部近邻约束的基于RGB-D的室内场景实时三维重建算法.该算法首先利用RGB-D相机采集到的RGB图像做哈里斯角点检测,再用SURF特征点描述方法对检测到的特征点生成64维特征描述子.接着利用特征点集合的深度信息和局部近邻特征点信息作为约束,初步筛选出相邻帧间正确的匹配点对,再结合随机抽样一致性(RANSAC)算法去除外点,以此得到相机的姿态估计.最后利用RGB-D的深度图像,在图优化方法(g2o)的基础上生成三维点云,实现室内场景的三维重建.实验中,RGB-D摄像头装载在自主移动导航的小车上,实时重构的三维场景验证了所提算法的可行性和准确性.

关键词: RGB-D, 三维重建, 特征点深度约束, 特征点局部近邻约束, 实时性

Abstract: A novel RGB-D based feature point depth constraint and locality constraint integrated real-time indoor scene 3D reconstruction algorithm is proposed, which focusing to solve problems such as the inaccuracy of 3D point cloud matching, excessive time consuming and the loss of point depth. Firstly, the feature points are detected using Harris point detector, and then labeled with 64 dimensional vector using SURF descriptor. Secondly, the initial correct feature point pairs are selected between the successive frames with the depth information constraint and feature point locality constraint in addition to vector similarity constraint. Thirdly, the outliers are removed and the camera pose is estimated based on the random sample consensus(RANSAC) method. Eventually, 3D point clouds are determined using the general graph optimization (g2o) facilitating the indoor scene reconstruction. In the experiments, the RGB-D camera is fixed on the automatic guided vehicle to capture the indoor surrounding scenes. Experimental results validate that the proposed approach is feasible and effective.

Key words: RGB-D, 3D reconstruction, feature point depth constraint, feature point locality constraint, instantaneity

中图分类号: