3D modeling of tree trunks based on LiDAR point cloud scans plays an essential role in forestry research, environmental protection, and ecological analysis. Due to factors such as uneven lighting conditions, overlapping objects, limited scanning positions, and point cloud registration errors, model reconstruction faces challenges with low data integrity. To overcome the limitations imposed by incomplete tree point cloud data on reconstruction, researchers often extract tree trunk skeletons based on geometric features of trees, then complete the tree point clouds along the skeletons, or directly proceed with model reconstruction. This dissertation studies three skeleton extraction methods including Laplacian-based contraction, $L_1$ medial skeleton and KNN-based contraction. We analyze the mathematical principles and algorithmic processes of these methods, after which they are implemented. Specifically, for the $L_1$ medial skeleton extraction method, a skeleton optimization algorithm is proposed, employing RANSAC-based local cylinder fitting and skeleton continuity detection. In the experimental section, the accuracy of the skeleton extraction algorithms implemented is compared, and the robustness of each algorithm is tested under simulated data with varying degrees of incompleteness, and real-world point cloud scans. Through qualitative analysis, the local and overall advantages and disadvantages of each algorithm in different scenarios are studied. Through quantitative analysis, a distance metric-based model for algorithm accuracy evaluation is established. The extraction accuracy of each method is evaluated using precise mathematical scales. The results show the extraction accuracy of each algorithm under the same simulated data at average and extreme incompleteness (60%): Laplacian—0.649/0.441, $L_1$—0.621/0.414, KNN—0.294/0.221, improved $L_1$—0.806/0.724. Combining qualitative and quantitative analysis, we conclude that among the classic methods, the $L_1$ medial skeleton extraction method has the best accuracy and robustness combined. Additionally, the skeleton optimization algorithm developed in this study improves the original $L_1$ method by an average of 29.79% in accuracy and by 74.88% under extreme conditions.