Aiming at the problems of low modeling accuracy, high cost and poor topology structure in the three-dimensional (3D) reconstruction of fruit trees, a 3D reconstruction method of fruit tree phenotype and skeleton extraction based on Kinect v2 sensor was proposed. Firstly, the Kinect v2 sensor was used to collect fruit tree point cloud data from different perspectives. Secondly, the characteristic point detection of scale invariant feature transformation was carried out on the plant point cloud, the eigenvector vector calculation was carried out by using the fast point feature histogram algorithm, the initial position of the point cloud was purified by the random sampling consistency method, and the improved iterative nearest point algorithm was used to finely register and stitch to form a complete point cloud after the initial transformation. Finally, the Delaunay triangulation of the point cloud data was used to fill the missing point cloud, the Dijkstra shortest path algorithm was used to obtain the minimum spanning tree, the skeleton was simplified by iteratively removing redundant components, the tree skeleton was estimated by the cylindrical fitting algorithm, and the tree skeleton was transformed into a closed convex polyhedron, and the 3D reconstruction of the branches of the fruit tree was realized. The experimental results showed that the average error of point cloud registration was 0.52cm, and the average error of branch reconstruction was not more than 3.52%, and the reconstruction effect was good. The research results can provide data support for orchard assessment of crop status, intelligent pruning and other research.