3D Reconstruction Of Apple Tree Based On Kinect

Three dimensional reconstruction of young apple tree for modeling of virtual trees has important significance. Professional 3D point cloud scanners have high accuracy, but they are expensive and too complex. Kinect has the advantages of small size, light quality and low cost, and these make it has been widely applied in 3D scanning and reconstruction of small objects. This research mainly focuses on the obtaining of 3D point clouds of fruit trees based on Kinect and reconstructing virtual model of the apple trees based on point clouds. To get three-dimensional information of trees timely, a cloud data capturing platform with Kinect was built. To realize the registration of multi-view point clouds, registration method based on spatial reference ball combined with ICP(Iterative Closest Point) algorithm was put forward.To get skeletons of apple trees, space colony algorithm was applied to extract skeletons. 3D models of apple trees were reconstructed, visualized and rendering based skeletons using pipeline model and general cylinder. In this research, 3~5 dwarf apple trees were tested on the experimental, and mainly research contents and conclusions were carried out as following.(1) 3D point cloud information capturing platform was built with Kinect.The experiment platform was built with a Kinect and a PC computer, and multi-view 3D point clouds were obtained successfully using it. Platform includes hardware(Kinect, PC computer and three tripod) and related API interface and support library equipped with PC computer. Kinect was fixed by a tripod support, its vertical height was at the objects central and changed tripod vertical and horizontal position to obtain better quality point cloud data. Experiments were performed under without direct sunlight and wind, the point clouds which were smoothed by down-sample met the requirements of point clouds registration in the next part.(2) The space reference spheres method combined with ICP algorithm of point cloud registration was put forward which realized the rough and accurate registration of multi-view point clouds of trees. According to the particularity of the tree point clouds, that is, tree point clouds are unable to estimate high quality vector, and rotary method or tagging method are not suitable to be applied of experiment condition, spatial reference spheres with the registration method of ICP algorithm as a novel registration method was proposed. Experimental put 3or 4space reference spheres around the trees when the point clouds were captured. Parameters of space reference spheres were calculated in different point cloud spaces that would be treated as coordinate transformation common points to obtain the transformation matrices. Then transformation matrices were applied to obtained rough registration point clouds. For theequipment precision leaded to that point clouds rough registration still existed obvious disparity, ICP algorithm was applied to point cloud registration which could make two pieces of point cloud reduce distance and become closer. Experimental results showed that the point cloud registration model length, width and height dimension error rate was controlled within10.00%, wherein the length and width dimensions error rate was controlled in less than 5.00%,the corresponding maximum error 0.0186 m, above results met the requirements of fruit tree skeleton extraction based on 3D point cloud in the next part.(3) The tree skeleton extraction and 3D model reconstruction were realized by using the spatial colony algorithm and pipeline model. Skeletons of the trees were extracted by space colony algorithm, the algorithm obtained skeleton points which were in good agreement with the model of point cloud through the iterative algorithm by setting and adjusting the influence radius, delete threshold, skeleton point distance and controlling angle. For the reconstruction of the fruit tree models, the experiment used the pipeline model and the generalized cylinder to reconstruct the branches. First, starting and ending of branches were established, and hierarchical structure of the trees were establish. Then, thickness of branches was controlled with the principles of the pipeline model, and branches were expressed by generalized cylinder. Finally, the leaves were added to the branches, and the tree models were visualized and rendered. Reconstruction results showed that the virtual models were similar with natural tree and had good visual effect. All expenrimental data and results verified that the proposed method was effective and reiable.