The current planting pattern of dwarf and dense jujube tree was increasingly conducive to mechanized harvesting, spraying and automated pruning, in which pruning of jujube tree removed redundant or overgrown branches and controlled tree structure to increase yield and extend life cycle. There were two main pruning methods: artificial pruning and whole geometric pruning. The quality of artificial pruning which was low efficiency, high labor cost and labor intensity was high. Higher efficiency can be achieved through whole geometric pruning with pruning machine adopted fixed distance, but wrong and missing pruning was serious. Automatic selective pruning reduced cost of labor, with a large number of useful branches being protected. In order to meet the requirement of automatic selective pruning of jujube tree, a complete 3D model of tree was needed. To ensure that two frames of point clouds of two fixed cameras were matched together with high accuracy, point cloud registration revisited algorithm was used. A system platform was built with two Azure Kinect DK depth cameras to obtain two fames of point clouds of a same tree in two positions with 55 degrees rotation difference. In order to register these two frames automatically, skeleton points based registration pipeline was proposed. Firstly, skeleton points’eigenvectors were calculated with fast point feature histograms (FPFH). Then, sample consensus initial alignment (SAC-IA) was applied for rough registering. Finally, fine registration with ICP algorithm was done with KD-tree acceleration to obtain a completed 3D branches model. The registration precision and time for the different natural environments were compared in the experiment, and results showed that illumination had a significant influence on the number of point clouds of jujube tree collected by the system. As a result, jujube branches 3D model were partially incomplete in sunny days, while complete jujube branches can be reconstructed in cloudy days and at night. In sunny days, the registration time was the minimum, only 0.09s, and error was the largest with fitting score of 0.00029;in cloudy day the registration time was 0.12s between sunny days and night, and the error was the smallest with fitting score of 0.00011. Compared with sunny and cloudy days, registration time was the longest at 0.16s at night, and the error was between sunny and cloudy days.