In order to avoid obstacles for fruit harvesting robot, the method of spatial information extraction of tree branch was studied. The method of normalized cross-correlation was used to get the disparity map of stereo image, branch skeleton was extracted and feature points were extracted by multi-segment approximation method. The spatial coordinates of branch feature points were calculated by combining with the disparity map based on binocular stereo vision, the radius of branch feature points were calculated by using the branch range image, in the end, the forked branch was divided into separate section. Rapid model reconstruction for fruit tree was researched: firstly, each branch module was constructed by 12-prism in the coordinate origin; then, every module was set to the correct posture by affine transformation; finally, the complete fruit tree model was restructured. Experiments show that the virtual trees provide the actual environment reference for obstacle avoidance and path planning of fruit picking machine.