With the rapid development of science and technology, precision agriculture shows more and more demands for 3D (three-dimension) reconstruction of crops, accordingly the demands for cheap sensor devices as well as high-efficiency 3D reconstruction solution have been becoming more and more urgent.With the development of data visualization technology, it becomes easier to measure the needed data for 3D reconstruction.Taking apple tree as the research object, this paper proposed an improved ICP (Iterative Closest Point) algorithm based on fruit tree three-dimensional point cloud registration method, and using OpenMP technology to achieve high-efficiency 3D reconstruction for apple tree.Using Kinect (Version 2, Windows crop.production) as 3D camera, two group images were taken from different perspective, and the color image and depth image of the apple tree were acquired as one image group.The image fusion functions which provided by Windows corp.were called to obtain three-dimensional point cloud data for apple tree by fusing its color and depth information from one perspective.After multi-source information fusion, high-efficiency reconstruction technology was explored to build accurate three-dimensional fruit tree quickly.Firstly, background removing and de-noising pretreatment was carried out for fruit tree's original point cloud and accordingly the relative accurate data were acquired as the fruit tree's point cloud.Secondly, using depth information of fruit tree's point cloud image and object boundary characteristics, the key points were extracted based on NARF (Normal Aligned Radial Feature) algorithm.This key points calculation method was based on depth image and not affected by external interference illumination variation.Thirdly,the FPFH (Fast Point Feature Histograms) algorithm was developed to get the key characteristic vector for each key point, and a compact representation of fruit tree point cloud was formed with rotation invariance and higher representative and descriptive.Fourthly, according to the key characteristic vectors, the correspondence points was estimated for two pieces of point cloud and correspondence errors could be eliminates using RANSAC (Random Sample Consensus) algorithm.And then, the corresponding relationship between two pieces of point cloud was obtained by applying the rotation matrix and translation vector through calculating the rotation parameters of the two pieces of point cloud, and accordingly rough registration was achieved according to the rotation parameters.Finally, on the basis of the coarse registration, the ICP algorithm was used to complete accurate registration for two pieces of point cloud.At the same time, aiming at improving registration process and decrease the cost of running time, an program was developed based on the acceleration of OpenMP mechanism, and the efficiency for point cloud registration processing was then guaranteed with unchanged precision and robustness.This paper carried out the experiment based on the above algorithms and the results showed that the improve point cloud registration method could be used to match two pieces of point cloud at any arbitrary initial positions, and its registration accuracy reached 0.0664 cm.And by using OpenMP technology, the efficiency of point cloud registration was significantly improved under the premise that the registration accuracy and stability were not affected.