基于点特征的图像—点云跨模态配准方法

A cross-modal registration method for image-point cloud based on geometric primitive features

随着互联网和多源传感器的快速发展,多源跨模态融合数据在测绘领域的应用不断扩大。点云数据和图像数据作为两种广泛使用的异源数据,各有优势,融合两者可实现传感器信息互补,获取更丰富的空间场景数据。配准作为关键环节,研究意义重大。本文在基于点特征的2D-3D配准思路下,对试验数据进行Harris角点特征提取,并手动匹配同名点对,评估了DLT(direct linear transform)、P3P和EPnP(efficient PnP)的配准效果。在配准的精度方面,试验组1中三者的RMSE为119.305、16.301和17.820像素;试验组2中的RMSE为14.031、20.322和9.858像素。在配准的效率方面,DLT算法配准耗时远高于其余两种算法。综合评估,EPnP算法性能优于前两者。通过高斯牛顿法对EPnP的位姿估计结果进行非线性优化,两组试验组的RMSE分别降至17.552像素和9.634像素,配准精度有所提升。

With the rapid development of the Internet and multi-source sensors,the application of multi-source cross-modal fusion data in the field of surveying and mapping is expanding.Point cloud data and image data,as two widely used heterogeneous data sources,each has their own advantages.By integrating the two,sensor information can be complemented,leading to richer spatial scene data.Registration,as a key process,holds significant research value.Based on a 2D-3D registration approach utilizing point features,this paper extracted Harris corner features from experimental data and manually matched corresponding points.The registration performance of DLT(direct linear transform),P3P,and EPnP(efficient PnP)is evaluated.In terms of the accuracy of registration,the RMSE for the three methods in experiment group 1 are 119.305,16.301,and 17.820 pixels,respectively;in experiment group 2,the RMSE are 14.031,20.322,and 9.858 pixels,respectively.Regarding the efficiency of registration,the DLT algorithm took significantly longer than the other two algorithms.Overall,EPnP performed better than the other two methods.The pose estimation results of EPnP are nonlinearly optimized using the Gauss-Newton method,reducing the RMSE of the two experimental groups to 17.552 pixels and 9.634 pixels,respectively,thus improving the registration accuracy.