The cultivation density of cherry trees affects the light distribution of their canopies. By studying the three-dimensional structure of group cherry trees, the light distribution of greenhouse sweet cherry trees under different planting densities can be analyzed, which can guide the scientific planting of cherry trees and improve the yield and quality of sweet cherry trees. High quality point cloud data was the basis of constructing the three-dimensional structure of the group cherry tree, and point cloud denoising and registration were the key steps of point cloud data preprocessing. A method for denoising and registration of group cherry trees based on 3D point cloud was proposed to build a 3D information acquisition platform for group cherry trees, and two fixed DK depth cameras were used to obtain the color point cloud data of group cherry trees. A binary classification method based on color region growth was proposed, and the color threshold was set to segment the point cloud and perform binary classification processing, which can effectively remove the abnormal invalid points in the color point cloud data, and set the dispersion of point cloud and RGB value as the evaluation standard of point cloud denoising. Combined with manual labeling method and dual camera pose matrix, an improved ICP method based on color features was proposed to solve the problem that traditional ICP registration algorithm depended on the initial pose and the registration speed was slow. SIFT algorithm was used to extract the color feature points, and the color feature points were combined with ICP algorithm for precise registration. Then the random sampling consistency algorithm in PCL was used to remove the wrong matching points, which effectively reduced the registration time and improved the registration accuracy. Taking 20 groups of group cherry tree point cloud data in summer and winter as experimental objects, the registration accuracy and registration time of ICP algorithm, NDT algorithm, SAC-IA algorithm and the proposed registration method were compared and analyzed. The results showed that the average registration time of the proposed registration method was 5.01s and 4.30s, respectively. The root mean square error was 2.316cm and 2.100cm respectively, which effectively reduced the registration time and registration error, and verified the effectiveness and universality of the proposed algorithm.