Wood Extraction from Urban Street Tree Point Clouds using a Deep Learning Approach Andrew Egbert Wiryawan (a), Chi-Kuei Wang (a*), Mohamad Bagas Setiawan (a), Michael Vashni Immanuel Ryadi (a)
(a) Department of Geomatics, National Cheng Kung University, Tainan City 70101, Taiwan R.O.C
Abstract
Accurate tree structure measurement is important for urban forest carbon assessment. For this purpose, Terrestrial Laser Scanner (TLS) enables these measurements through 3D point clouds that capture detailed tree structure. However, extracting wood components from point clouds has remained challenging due to incomplete data from occlusions and species diversity. Recent advances in deep learning have shown promising results for wood extraction. Most of the proposed deep learning techniques for extracting the wood point clouds use point-based methods, voxel-based methods, and projection-based methods. Among these three methods, the projection-based method has its potential for further exploration. Therefore, this study proposes a workflow that projects 3D voxelised point clouds into multiview 2D cross-sectional images. Our method employs Swin-UPerNet networks on the 2D cross-sectional images to perform wood extraction in urban tree point clouds. We apply average voting to determine the final predicted class of point clouds. This method examines three different urban scenario datasets, i.e. one virtual dataset and two real-world datasets. Our results show a promising accuracy of wood point clouds. Furthermore, the findings demonstrate the potential of projection-based approaches in tackling various species and incomplete data due to occlusion challenges for extracting the wood structure from urban forest scenes.
