激光点云与密集匹配点云融合方法

Fusion Method of LiDAR Point Cloud and Dense Matching Point Cloud

针对地面激光扫描点云和航摄影像密集匹配点云融合存在质量退化与分层的问题,本文提出了一种基于图割算法和引导点云滤波算法的异源点云融合方法。该方法首先利用一种结合几何和颜色信息的图割算法分割密集匹配点云,然后利用分割后的密集匹配点云填补激光点云的孔洞和遮挡区域,接着采用以表面曲率加权的引导点云滤波算法消除混合边界处的缝隙并纠正混合点云中的平面错层。实验结果证明,所提方法对比现有方法有明显的性能提升,融合后的精度和完整性分别提升了5.42%和2.94%,能够较好地支撑激光点云与密集匹配点云的高质量融合。

Objective The fusion of three-dimensional(3D)laser scanning and photogrammetry is a hot topic in urban 3D reconstruction.Using the dense point cloud obtained from the oblique images to improve the integrity of the scene entity,overcome the limitation of laser scanning technology,which is difficult to scan completely,and realise the fusion of different source point clouds with varying accuracy has significant research implications.The reconstructed method by fusing multi-source point clouds can take advantage of the complementary characteristics of multi-source point clouds and is one of the effective methods for improving model reconstruction quality and solving the data loss problem.However,owing to acquisition methods and equipment errors,there are many differences in multi-source point cloud data,such as point cloud density,distribution uniformity,accuracy,noise,coverage and occlusion.The current standard method is multi-source point cloud registration.On the one hand,the aligned point clouds improve the reconstructed completeness of the scene,but there are bound to be stratified,redundant noise in the mixed point clouds,and the aligned point cloud’s edge is not smooth.On the other hand,point clouds with noise and redundant information will directly degrade the quality of surface reconstruction.We propose a fusion method based on a graphcuts algorithm and guided 3D point cloud filtering algorithm to address the problems of quality degradation and stratification in the fusion of terrestrial laser scanning(TLS)point clouds and dense matching point clouds from aerial oblique images.Methods In this paper,a new fusion method of TLS cloud and the dense matching point cloud is proposed.First,dense matching point clouds are split using the graph-cuts algorithm,which combines geometry and colour information.The data term of the energy function is built based on the distance between the multi-view stereo(MVS)point cloud and the TLS point cloud,as well as the angle between their normal vectors,and the geometric neighbourhood relationship and colour difference in the MVS point cloud are then combined to build the smooth item.The graph-cuts algorithm is used to optimise the MVS point cloud’s binary classification label sets,and the overlapping redundant area of the MVS point cloud is removed based on the TLS point cloud.The remaining dense point cloud is then used to fill the gaps and occlusions in the LiDAR point cloud.This paper proposes a neighbourhood point selection strategy that selects an appropriate proportion of dense matching points and LiDAR points as guided filtering neighbourhood point sets for the points to be processed that are near the boundary of the LiDAR point cloud.Finally,the guided 3D point cloud filtering algorithm weighted by surface curvature is used to reduce blending boundary gaps and correct stratified redundancy in the blended point cloud.Results and Discussions This paper conducts qualitative and quantitative experiments to demonstrate the efficacy of the proposed method in improving the quality of blended point clouds.The three data sets are described in detail in Table 1.The first type of data is a dense matching point cloud and ground laser scanning point cloud obtained from aerial oblique images of a laboratory building(Fig.4).The proposed method eliminates the gap in the overlapping area of TLS point cloud and MVS point cloud,reduces noise at the left edge of the building,and corrects and integrates the stratified layer of MVS point cloud into the TLS point cloud of the building facade,resulting in multisource point cloud fusion with different accuracy(Fig.5).When the graph-cuts algorithm fails to remove the overlapping area of the MVS point cloud ideally,the guided point cloud filtering algorithm weighted by the surface curvature can move the dense point cloud as a whole to the plane of the LiDAR point cloud,removing redundancy and stratified layers(Fig.6).When compared to the progressive migration algorithm,the proposed method improves the accuracy and completeness of the dense point cloud smoothed.The accuracy of the smoothed MVS point cloud of Courthouse data has increased from 67.51%to 71.17%,and the completeness of the Courthouse data has increased from 77.96%to 80.25%(Table 3).The guided point cloud filtering algorithm weighted by the surface curvature can adjust the smoothing parameter adaptively based on the flatness of the point cloud neighbourhood,improving the smoothing effect without destroying the original point cloud structure.Experiments show that the proposed method improves the accuracy and completeness of the fused dense matching point cloud,and they are improved by 5.42%and 2.94%,respectively,when compared to existing methods.The proposed method can help with a high-quality fusion of LiDAR point cloud and dense matching point cloud.Conclusions This paper proposes a TLS point cloud and MVS point cloud fusion method based on the graph-cuts algorithm and the guided point cloud filtering algorithm.The graph-cuts algorithm is used to fuse the geometry and colour information of heterogeneous point clouds,and the neighbourhood relationship is considered to ensure the consistency of dense matching point clouds after segmentation.Then,using a neighbourhood point selection strategy,an appropriate proportion of the guide point cloud filtering neighbourhood points are chosen for the dense point cloud near the boundary of the LiDAR point cloud,to realise the guided point cloud filtering weighted by the surface curvature.The experimental results of fusing heterogeneous point clouds with different accuracy,smoothing the gap at the junction of mixed point clouds,and correcting stratification are realised.Experiments show that the proposed method can effectively remove the overlapping and redundant parts of the MVS point cloud and the LiDAR point cloud,as well as improve the accuracy and completeness of the dense matching point cloud,which benefits surface reconstruction.