Rapid 3D reconstruction of fruit tree based on point cloud registration

In order to provide 3D visual image of fruit tree in real-time for guiding orchard production management, a method of 3D reconstruction of fruit tree based on point cloud registration was proposed in this paper.The color images for their corresponding depth of fruit tree were taken by using RGB-D camera from multiple aspect angles.Then 3D point clouds of fruit tree in different perspectives were computed and acquired through fusing the corresponding information of its color image and depth image.And a high-efficiency point cloud registration approach was explored and tested to reconstruct 3D fruit tree’s point cloud model quickly: Firstly, data preprocessing of fruit tree’s each piece of point cloud was carried out for background removing and original point cloud de-noising based on depth distance judgment and spare noise filtering methods respectively, and accordingly each relative accurate data set was obtained as the fruit tree’s point cloud in its each specific perspective.Secondly, the key points of each piece of point cloud were extracted using Normal Aligned Radial Feature (NARF) algorithm based on the depth and boundary information of fruit tree’s point cloud, and their corresponding feature vectors were also calculated using Fast Point Feature Histograms (FPFH) descriptor.Thirdly, the feature vectors were compared between two adjacent pieces of point cloud and then pairs of corresponding key points were estimated and extracted.Then those pairs of corresponding key points were validated and refined using the RANdomSAmple Consensus (RANSAC) algorithm to obtain the correct space mapping relationship between two adjacent pieces of point cloud, and further the transformation parameters from one piece of point cloud to its adjacent one were computed.And then, the initial registration of two adjacent pieces of point cloud was completed by transforming them to the same coordinate system according to their transformation parameters.Fourthly, on the basis of the initial registration, the Iterative Closest Point (ICP) algorithm was implemented to achieve accurate registration for two adjacent pieces of point cloud.Finally, using the initial and precise registration algorithm mentioned above, the remaining pieces of point cloud were globally matched and 3D reconstruction of whole individual fruit tree’s point cloud model was then realized.Moreover, aiming at decreasing the cost of running time of point cloud registration, a program was developed based on the acceleration of OpenMP mechanism, and the efficiency for point cloud registration process was significantly improved with precision and robustness unchanged.The experiment was carried out and the results showed that the proposed approach could be used to match pieces of point cloud at any arbitrary initial positions to reconstruct 3D point cloud for fruit tree rapidly, and its registration distance error was 0.0068 m.