• Optics and Precision Engineering
  • Vol. 30, Issue 22, 2901 (2022)
Jia-le ZHOU, Bing ZHU*, and Zhi-lu WU
Author Affiliations
  • College of Electronic and Information Engineering, Harbin Institute of Technology, Harbin150001, China
  • show less
    DOI: 10.37188/OPE.20223022.2901 Cite this Article
    Jia-le ZHOU, Bing ZHU, Zhi-lu WU. Camera pose estimation based on 2D image and 3D point cloud fusion[J]. Optics and Precision Engineering, 2022, 30(22): 2901 Copy Citation Text show less
    Relationship between coordinate systems
    Fig. 1. Relationship between coordinate systems
    3D sparse point cloud reconstruction and pose dataset construction
    Fig. 2. 3D sparse point cloud reconstruction and pose dataset construction
    Dense point cloud data construction
    Fig. 3. Dense point cloud data construction
    Flow of multistage camera pose estimation
    Fig. 4. Flow of multistage camera pose estimation
    Flow of RANSAC algorithm
    Fig. 5. Flow of RANSAC algorithm
    Structure dense scene regression network
    Fig. 6. Structure dense scene regression network
    Result of scene regression
    Fig. 7. Result of scene regression
    Accuracy of pose estimation varies with the error threshold
    Fig. 8. Accuracy of pose estimation varies with the error threshold
    Pose estimation accuracy comparison(5 cm/5°)
    Fig. 9. Pose estimation accuracy comparison(5 cm/5°)
    Results of median localization Errors
    Fig. 10. Results of median localization Errors
    数据集名称Labcore
    RGB640×480
    Depth640×480
    深度值(16位)单位mm
    帧数4 000
    帧率30 fps
    训练样本数2 000
    测试样本数2 000
    Table 1. Detail of labcore dataset

    数据集

    场景

    数据集组成

    Train Test

    位姿估计方法
    PoseNet5DSAC++15SANet20Pixloc16多阶段场景回归(本文)
    Chess4 000 2 00032.0 cm/4.06°2.0 cm/0.5°3.0 cm/0.88°2.0 cm/0.80°1.8 cm/0.64°
    Fire2 000 2 00047.0 cm/7.33°2.0 cm/0.9°3.0 cm/1.08°2.0 cm/0.73°1.8 cm/0.86°
    Heads1 000 1 00029.0 cm/6.00°1.0 cm/0.8°2.0 cm/1.48°1.0 cm/0.82°1.2 cm/0.72°
    Office6 000 4 00048.0 cm/3.84°3.0 cm/0.7°3.0 cm/1.00°3.0 cm/0.82°2.7 cm/0.81°
    Pumpkin4 000 2 00047.0 cm/4.21°4.0 cm/1.1°5.0 cm/1.32°4.0 cm/1.21°4.0 cm/1.11°
    RedKitchen7 000 5 00059.0 cm/4.32°4.0 cm/1.1°4.0 cm/1.40°3.0 cm/1.20°4.0 cm/1.10°
    Stairs2 000 1 00047.0 cm/6.93°9.0 cm/2.6°16.0 cm/4.59°5.0 cm/1.30°3.5 cm/1.01°
    Table 2. Performance comparison of camera pose estimation algorithms
    处理步骤DSM9DSAC++15Pixloc16本文
    图像检索(单帧)0.170.170.17
    位姿估计(单帧)0.210.20.680.18
    总耗时(单帧)0.380.20.850.35
    Table 3. Time consumption comparison of camera pose estimation algorithms
    Jia-le ZHOU, Bing ZHU, Zhi-lu WU. Camera pose estimation based on 2D image and 3D point cloud fusion[J]. Optics and Precision Engineering, 2022, 30(22): 2901
    Download Citation