Geometric and Multi-Scale Feature Fusion for Complete Tree Skeleton Extraction

Highlights A tree skeleton extraction method based on the fusion of geometric and multiscale features. DBSCAN clustering resolves skeleton omissions and merging in graph algorithms for incomplete point clouds and adjacent branches. Unique breakpoint selection strategy adaptable to various breakage scenarios. Combining local and global features to enhance the accuracy of skeleton breakpoint connections. ABSTRACT. Tree topology reconstruction is essential for applications in precision agriculture, such as canopy structure analysis and yield prediction. However, 3D reconstruction based on RGB images is often affected by noise and data loss, exacerbating the problem of missing branches during skeleton extraction. To address this issue, this paper proposes a tree skeleton extraction method that integrates local and global geometric features. First, DBSCAN is introduced into the graph-based clustering process to segment the incomplete point cloud into multiple clusters, facilitating skeleton extraction in discontinuous regions and enhancing the distinction of spatially adjacent branches. Then, potential connections between skeleton segments are identified based on local branch distance and angular features. Incorrect connections are removed through closed-loop structure recognition and filtering, ensuring accurate skeleton completion. Furthermore, the global growth direction of the tree is incorporated to refine the skeleton structure, followed by Laplacian smoothing to enhance skeleton quality. Experimental validation on point clouds from 13 real plum trees and 200 simulated trees demonstrates the effectiveness of the proposed method, achieving an accuracy of 85.96% for real trees and 88.80% for simulated trees. The results significantly improve the completeness and accuracy of tree structure reconstruction, providing a reliable approach for subsequent tree topology analysis and precision agriculture applications.