基于RGB-D相机的球体三维重建及定位方法研究  

3D reconstruction and positioning for spheres based on RGB-D camera

作　　者：冯春[1] 花省 王树磊 尹飞鸿[1] 江炜[1] FENG Chun;HUA Xing;WANG Shu-lei;YIN Fei-hong;JIANG Wei(Department of Mechanical and Vehicular Engineering, Changzhou Institute of Technology, Changzhou 213032, China)

机构地区：[1]常州工学院机械与车辆工程学院,江苏常州213032

出　　处：《机电工程》2019年第5期549-554,共6页Journal of Mechanical & Electrical Engineering

基　　金：江苏省自然科学基金青年项目(BK20140252;BK20150256)

摘　　要：针对单个和多个球体识别和定位问题,采用一种利用RGB-D相机进行三维重建的方法,对RGB-D相机的标定、深度图像和彩色图像的配准、彩色三维点云的生成等进行了研究。提出了一种阈值分割结合随机采样一致性算法的定位方法,利用色彩阈值分割和深度阈值分割方法实现了背景分割,得到了整体的球体区域的点云数据;利用欧几里得点云分割方法将整体的球体区域分割成了独立的球体点云,并辅以基于曲面法线的阈值分割方法分离相接触球体的点云;采用随机采样一致性算法分别计算出了每个球体的参数,得到了球体的球心坐标、半径以及深度信息。研究结果表明:在不同光照条件下该方法具有较好的实时性和定位精度,在0.5 m～1.5 m的深度范围内,半径30 mm球体的半径平均测量误差为4 mm,球心坐标在x,y,z方向的平均偏差为1.3 mm、2.6 mm、1.4 mm,单帧图像处理平均时间约为1 s。Aiming at single and multiple sphere recognition and localization problems, a method of three-dimensional reconstruction using an RGB-D camera was employed. The calibration of RGB-D camera, the registration of depth image and color image, and the generation of color 3 D point cloud were studied. A positioning method based on threshold segmentation combined with random sampling consistency algorithm was proposed. The method of color threshold segmentation and the depth threshold segmentation method was used to realize the background segmentation, and the point cloud data of the whole sphere region was obtained. The Euclidean point cloud segmentation method was applied to divide the whole sphere region into independent sphere point clouds, and the threshold segmentation method based on surface normal was used for separating the point cloud of the contact sphere. The random sampling consistency algorithm was used to calculate the parameters of each sphere separately, and get the spherical coordinates, radius and depth information of the sphere. The results indicate that the method has good real-time and positioning accuracy under different illumination conditions. In the range of 0.5 m^1.5 m depth, the average radius error of the sphere with a radius of 30 mm is 4 mm, and the spherical coordinates error at x, y and z direction is 2.3 mm, 4.5 mm, 3.7 mm in average. Meanwhile single frame image average processing time is about 1 s.

关 键 词：RGB-D相机 三维重建 识别 定位 点云分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]