Accurate and efficient point cloud branch and foliage separation are essential for accurately calculating above-ground biomass and carbon stocks. However, current methods are computationally expensive and rely on priori knowledge leading to insufficient generalization. To address the above problems, the article proposes to use the point feature-based Transformer network for automated branch and leaf separation study in forest scene 3D point cloud. This research uses the Point Transformer v2 network. Firstly, the grid coding module is used to extract the learnable local structural relations and preserve the geometric topology of the point cloud; secondly, group attention is used to achieve multi-channel joint learning, reduce the redundancy of features and improve the efficiency of computation; finally, a point-based Transformer network is constructed to achieve high-precision semantic segmentation of 3D point clouds of forest trees, and this method is capable of accurate separation of tree trunks and leaves. In this paper, we use the 3D point cloud data of seven different tree species sample plots in Canada and Finland acquired by ground-based laser scanner to conduct branch and foliage separation experiments and accuracy evaluation, and the experimental results show that the OA of the network proposed in this paper is 94.42%, and the mIoU is 78.89%, which indicates that the network in this paper can effectively improve the segmentation accuracy of the tree crown under the conditions of irregular forest tree crown distribution and occlusion. Meanwhile, more accurate segmentation can be achieved for branch and foliage of different tree species and different point cloud densities.