关键词: 树木点云配准, 快速点特征直方图, Kinect相机, 采样一致性初始配准, 最近点迭代

Abstract: With the development of intelligence and informatization of agriculture and forestry, the 3D reconstruction technology of trees has gradually become the focus of research at home and abroad. In order to obtain a 3D morphological model of tree with color information, using simulation to simulate fruit tree, a 3D point cloud registration scheme of tree in two perspectives based on FPFH (Fast Point Feature Histogram) was proposed. Firstly, the calibration object was set above the target tree. The RGB image and depth image of the trees were obtained simultaneously from different angles using the Kinect camera. With the help of Microsofts official development tool Kinect for Windows SDK, the two images were aligned and fused. Then the initial point cloud with color information was generated by PCL (Point cloud library). Secondly, the point cloud of tree and calibration objects was extracted from XYZRGB initial point cloud. Considering the accuracy of acquisition equipment and measurement errors, there were some points far away from the main body in the target point cloud, which would to be eliminated. In order to improve the processing speed of the following point clouds, voxelated mesh sampling was applied to the point cloud. Thirdly, the FPFH of each point on the tree and the calibrator was calculated. According to FPFH, the corresponding points were searched, and initial registration was carried out by SACIA (Sample Consensusinitial Alignment). Thus, an approximate transformation matrix was estimated so that two pieces of point cloud got a better initial position. Finally, ICP (Iterative Closest Point) algorithm was used for further registration. In addition, the influence of Kinect acquisition distance, height, and angle difference on registration accuracy was discussed, and the best data acquisition method was determined: the distance from the camera to the center of the tree was 2 m, the height of the camera was 10 cm, and the interval of each scan was 45°. The experiment showed that setting the calibrator can acquire more discriminant FPFH features, which increase the accuracy of corresponding point query in initial registration and then improve the quality of registration, and the final registration error was less than 1.9 cm.

Key words:  registration of tree point cloud, fast point feature histogram, Kinect camera, sample consensusinitial alignment, iterative closest point