The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 217–226, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-217-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 217–226, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-217-2020

  12 Aug 2020

12 Aug 2020

WEIGHTED ICP POINT CLOUDS REGISTRATION BY SEGMENTATION BASED ON EIGENFEATURES CLUSTERING

M. Franzini1, A. M. Manzino2, and V. Casella1 M. Franzini et al.
  • 1Dept. of Civil Engineering and Architecture, University of Pavia, Italy
  • 2Dept. of Environment, Land and Infrastructure Engineering, Polytechnic of Turin, Italy

Keywords: UAV, Photogrammetry, Dense Point Cloud, Registration, Eigenfeatures, K-Means, Clustering, ICP

Abstract. Dense point clouds can be nowadays considered the main product of UAV (Unmanned Aerial Vehicle) photogrammetric processing and clouds registration is still a key aspect in case of blocks acquired apart. In the paper some overlapping datasets, acquired with a multispectral Parrot Sequoia camera above some rice fields, are analysed in a single block approach. Since the sensors is equipped with a navigation-grade sensor, the georeferencing information is affected by large errors and the so obtained dense point clouds are significantly far apart: to register them the Iterative Closes Point (ICP) technique is applied. ICP convergence is fundamentally based on the correct selection of the points to be coupled, and the paper proposes an innovative procedure in which a double density points subset is selected in relation to terrain characteristics. This approach reduces the complexity of the calculation and avoids that flat terrain parts, where most of the original points, are de-facto overweighed. Starting from the original dense cloud, eigenfeatures are extracted for each point and clustering is then performed to group them in two classes connected to terrain geometry, flat terrain or not; two metrics are adopted and compared for k-means clustering, Euclidean and City Block. Segmentation results are evaluated visually and by comparison with manually performed classification; ICP are then performed and the quality of registration is assessed too. The presented results show how the proposed procedure seem capable to register clouds even far apart with a good overall accuracy.