Segmentation of individual trees in urban MLS point clouds using a deep learning framework based on cylindrical convolution network

Automatic and accurate instance segmentation of street trees from point clouds is a fundamental task in urban green space research. Previous studies have achieved satisfactory tree segmentation results in simple scenarios. However, for challenging cases, including adjacent overlapping tree crowns, irregular tree shapes, and incompleteness caused by occlusion, most methods show under- or over-segmentation effects. In this study, an automated two-stage framework (tree extraction and individual tree segmentation) using vehicle-mounted mobile laser scanning (MLS) point clouds is developed to robustly detect single roadside trees. In the first stage, the ground points are filtered to reduce the processing time. Subsequently, an improved graph-based semantic segmentation network extracts roadside tree points from the urban scenes. For individual tree segmentation, a segmentation strategy combining cylindrical convolution and dynamic shift detects instance-level roadside trees. A simple road environment and two complex urban areas are used to verify the performance of the individual urban tree segmentation. The proposed method achieves 84-92% overall segmentation accuracy of the roadside tree point clouds and significantly outperforms existing methods in various challenging cases. Some applications can benefit from individual tree segmentation. For instance, the 3D green volume is calculated at the level of individual urban trees. The proposed method provides a practical solution for ecological assessment based on the 3D green volume of urban roads.