High precision 3D reconstruction and target location based on the fusion of visual features and point cloud registration

• Published in: Measurement : journal of the International Measurement Confederation 2025-02, Vol.243, p.116455, Article 116455
• Main Authors: Chen, Junliang, Wei, Xiaolong, Liang, Xiaoqing, Xu, Haojun, Zhou, Liucheng, He, Weifeng, Ma, Yunpeng, Yin, Yizhen
• Format: Article
• Language: English
• Subjects: Image retrieval; Pose estimation; SLAM; Three-dimensional reconstruction
• ISSN: 0263-2241
• DOI: 10.1016/j.measurement.2024.116455

•Novel 3D reconstruction method for large targets like aircrafts.•Utilizes ORB-SLAM3 for rapid camera pose estimation and sparse mapping.•Integrates image feature retrieval and point cloud registration for localization.•Introduces visual feature integration to strengthen point cloud registration.•Achieves significant performance improvements in trajectory estimation accuracy. Aiming at the requirements of large amount of data, complex calculation, high real-time and accuracy in the accurate 3D reconstruction of large targets such as aircrafts, a 3D positioning and reconstruction method based on the fusion of visual features and point cloud registration is proposed in this paper. Firstly, ORB-SLAM3 is used to realize fast camera pose estimation and sparse 3D map construction for large targets such as aircrafts. At the same time, image feature retrieval and matching technology is used to provide a reliable basis for coarse perception positioning of targets. Then, using the point cloud data of the large target obtained by the outdoor structured light camera, combined with the local point cloud of the model, through the point cloud registration technology, the accurate position estimation of the aviation large target is achieved. In order to further improve the accuracy and robustness of registration, this paper also integrates visual feature information, and provides more constraints for the registration process by matching feature points and associating them with point cloud data. The method proposed in this paper realizes the rapid coarse sensing positioning of large targets such as aircrafts, and realizes the accurate position estimation of targets through the point cloud registration technology, and finally completes the high-quality three-dimensional reconstruction. Through comparison and ablation experiments, the trajectory estimation error of the proposed method is 1.251 m, which is improved by 6.7 % and 23.2 % compared with ORB-SLAM3 and COLMAP, and the average MAE is reduced by 29.6 %, which has made a great improvement in accuracy.