Research On 3D Point Cloud Acquisition And Reconstruction Mehtod Of Plant Based On Kinect

With the development of digital agriculture,3D modeling technology of plants is applied more and more widely in agricultural production. Studying on the 3D structure of plants has a guiding significance for the research of crops planting spacing, intercropping mode, pruning method, drug spraying and crop yield changes.3D scanner was usually used to obtain the 3D data in plant 3D reconstruction, which is expensive and the data acquisition process was complex.In this paper, corn and other plants were used as the research object. Kinect was used to obtain the plants’ 3D point cloud data. Several parts were researched such as point cloud data preprocessing, denoising method, high effective registration method for 3D point cloud data of different positions as well as the method of reconstructing 3D plant model from the point cloud data, which was convenient for the later rendering. The main research work and conclusions are as follows:(1) Analyze and determine the method of obtaining the 3D data of plants. The corn and other plants were used as the main research object. The Kinect was used to obtain plants’ depth data from eight main angle and four auxiliary angle.10 frame data were obtained from each view and transformed into the 3D point cloud data with color information.(2) Point cloud data preprocessing. The through filtering was used to remove the background data. The multi frame data fusion method was used to get a more complete point cloud data and it played a role in preliminary smoothing. A point cloud denoising method based on the combination of statistical analysis and depth data bilateral filtering was proposed in this paper, which was used to deal with the outliers and high frequency noise. Tests showed that this method could well remove the noise and improve the quality of point cloud data obtained from Kinect. The internal high-frequency denosing time of corn and Spathiphyllum of algorithm in this paper was only 2.71% and 1.78%. of the traditional bilateral filtering algorithm.(3) A point cloud registration method based on calibration objects was used in this paper. Eight round slices with 70mm diameter were placed around the reconstruction object as calibration objects when acquiring the data. The transformation and rotation matrix of adjacent views was calculated by using the common calibration objects from different perspectives. The ICP algorithm was used for data accurate registration. Finally, the complete 3D point cloud model of the plant was obtained by matching the data of each angle to the same coordinate system.(4) In this paper, a new sampling method based on density of point cloud was proposed. The point data was simplified according to the preset sampling density. This method could simplify the point cloud data quickly and effectively, the neighborhood clustering method was used to optimize the data distribution. Finally, the greedy projection triangle mesh algorithm based on PCL was used for grid reconstruction. To present the natural color of plant the initial color of data is mapped to the point cloud.