Automated registration of forest point clouds from terrestrial and drone platforms using structural features

Abstract

Light Detection and Ranging (LiDAR) technology has demonstrated significant effectiveness in forest remote sensing. Terrestrial Laser Scanning (TLS) and Drone Laser Scanning (DLS) systems reconstruct forest point clouds from distinct perspectives. However, a single-platform point cloud is insufficient for a comprehensive reconstruction of multi-layered forest structures. Therefore, registration of point clouds from multiple platforms is an important procedure for providing comprehensive three dimensional reconstruction of the trees for more accurate characterization in forest inventories. However, the irregular and intricate structures of forest scenes, which often lack easily recognizable geometric features such as lines and planes, present substantial challenges for existing registration algorithms, such as Coherent Point Drift(CPD), Fast Global Registration(FGR), and Four Points Congruent Sets(4PCS). To address these challenges, we develop a novel algorithm, namely ForAlign, for the registration of forest point clouds from TLS and DLS. Our algorithm incorporates a tree location-based matching procedure followed by dynamic programming for detailed alignment. It fully considers the issue of inconsistent point cloud density distributions from different platforms and utilizes differential entropy to identify subsets of points with consistent structural features from the two data sources. These subsets serve as the basis for point cloud alignment based on distribution information. To validate the generality and accuracy of the proposed ForAlign, we conducted experiments using both scanned and simulated data describing different forest environments. The results show that our method achieves superior performance, with an average translation error of 6.4 cm and a rotation error of 53.5 mrad, outperforming CPD, FGR, and 4PCS by 43.5%, 55.4%, and 44.0% in translation accuracy, and by 36.4%, 54.6%, and 42.4% in rotation accuracy, respectively. Our study demonstrates that ForAlign effectively mitigates the errors introduced by tree localization in the preprocessing steps caused by varying point densities in TLS and DLS datasets, successfully extracts corresponding tree features among complicated forest scenes, and enables a robust, automated end-to-end registration process. The source code of ForAlign and the dataset are available at https://github.com/yiliutan/ForAlign.