Point cloud obtained from terrestrial laser scanner contains detailed, high precision three-dimensional(3D) surface coordinates, which is of special importance for forest parameter estimation and accurate reconstruction of plant model. An improved method for tree branching structure reconstruction was proposed based on the fact that a clean partition of branches and leaves form tree point cloud was very difficult if it was not impossible. Firstly, principal direction at each point was estimated with chord and normal vectors (CAN), and point cloud from the branches and leaves was separated by using both the similarity of principal direction between neighboring points and distribution density of points. Secondly, skeleton nodes and corresponding radii were computed from main branches by using level sets and least square method. For the leaves, the crown volume was divided into equal-sized voxels, all the points in a voxel were represented by the voxel’s centroid, and all centroid points formed feature points of the crown. Finally, tree model was reconstructed by cylinder fitting based on the topology of skeleton nodes and feature points. Segmentation results accuracy analysis and four different tree species model reconstruction examples were introduced. Segmentation accuracy analysis and model reconstruction quality evaluation showed that the approach was robust and insensitive to noise; the reconstructed tree models were in good agreement with the point cloud. The method was also able to extract structural parameters, including tree height, diameter at breast height (DBH) and volume parameters.