Accurate tree skeleton reconstruction and growth parameter retrieval from LiDAR data is a necessity tendency for precision forestry. For the scanned tree points, a lattice projection strategy fused deep learning network was proposed to realize wood-leaf point segmentation, and then computer graphics techniques were employed to realize tree skeleton modeling based on the segmented branch points. The proposed deep learning network includes the rotation invariance module, the lattice projection and barycentric interpolation module, the multi-scale transforms and convolution layers. Through the projection of the point clouds onto three coordinate planes according to lattice projection strategy and barycentric interpolation with the aim of appropriate network coefficient retrieval, the problem that 3D point clouds with the unordered spatial arrangement leading to difficulties in performing convolution operations is solved. Validated by many tree species in Hainan Island, the effectiveness of our approach was verified. The specific steps are as follows. First, the training sets composed of the scanned tree points with labeled wood-leaf class information was fed into the constructed deep learning network to optimize the weights of neuron networks. Then, the trained network was employed to segment wood-leaf points of the test samples composed of various tree species. Second, the segmented branch points were stratified into many layers according to a fixed height interval and the cluster centers of each layer were extracted using a spatial clustering algorithm. The connection relationship of the cluster centers belonging to each adjacent layer was determined based on the minimal Dijkstra distance and random sample consensus (RANSAC) was adopted to calculate the radius of each cylinder fitting every branch segment to constitute the whole tree skeleton. Finally, according to the linked list consisted of the central points of each layer and the criterion of minimum angle change, the trunk and first-order branches of the tree skeleton were automatically identified, which was utilized to assess the foliage clump volume borne on the corresponding branches. Compared with the manually leaf-wood segment results, the accuracy of classification using deep learning network is 91.31%, which is 7% higher than machine learning algorithms. A sound performance is also achieved for the estimated diameter of the first-order branches (coefficient of determination (R2) =0.92, root mean square error (RMSE) =0.64cm, relative root mean square error (rRMSE) =6.39%) and the included angle between first-order branches and trunk (R2=0.91, RMSE=3.32°, rRMSE=7.81%) using our method versus the field measurements for the study rubber trees. The proposed approach employed computer graphics theories and deep learning techniques to afford automatic and accurate tree skeleton reconstruction methodologies from raw scanned points, which have the potential and applicability in a range of tree species.