LAUSR

Street trees extraction and further acquisition of their property information is one of the current research hotspots, which can be applied in various urban management. For the processing task of automated identification of tree points from large-scale point clouds in urban road scenes, we propose an improved RandLA-Net network that takes into account both extraction accuracy and complexity. The method can eliminate the problem of insufficient feature extraction and more redundancy within the RandLA-Net model, and achieve efficient recognition of the street trees. As an application of street tree recognition results, we studied the accurate calculation method of street tree shading area and designed a density-based iterative $\alpha $ -shape algorithm. The method effectively solves the problems of large density differences among different point sets and overestimation of shading area caused by the existing convex hull algorithms’ unrefined contour reconstruction, and improves the calculation accuracy of the canopy shading area. Experiments on the Paris-Lille-3D dataset show that the improved RandLA-Net network improves the IoU of three classes by 3.19%~6.46% and reduces the model complexity by 1.4% compared to the original model. The density-based iterative $\alpha $ -shape algorithm achieves more refined contour reconstruction and solves the overestimation of shading area by 26% and 9% due to the other two convex hull algorithms. Its ability to effectively calculate accurate shading area was confirmed.