点云孔洞定位_基于TSDF模型的点云孔洞修复方法

摘要

深度摄像头的点云集合一般存在黑色孔洞闪烁的现象,在识别机械臂静态目标抓取时,点云集合数据的关键位置的体素不规则出现,就会对目标中心点的识别造成无法修正的偏差。因此,引入多个点云集合累积的思想,提出了一种基于TSDF模型的点云孔洞修复算法。在图形处理器(Graphics Processing Unit,GPU)中构建长宽高都是512体素的截断符号距离函数(Truncated Signed Distance Function,TSDF)模型,给定深度摄像头相对于世界坐标系的初始位姿,便可以得到相机坐标系到世界坐标系的坐标变换矩阵。根据当前图像坐标点的深度值和模型体素值,动态计算权值,更新模型,经过多次迭代,从而形成稳定的点云坐标集合。实验结果表明,所提出的算法应用在机械臂静态目标抓取后,被识别目标中心点重定位精度误差在2.0 mm内,机械臂抓取目标的成功率显著提高。由于模型在GPU中构建,并不会降低计算机工作性能。所提算法在修复方面的可靠性强于常规修复算法,对于颜色复杂的物体和对于没有彩色图像的深度摄像头依然适用。

The phenomenon of black hole flashing always exists in point cloud set of the depth camera.The irregular appearance of the data voxel of the key position will bring uncorrectable bias to target center identification during recognizing robotic static target grapping. A new algorithm for point cloud hole repairing algorithm based on TSDF model is proposed through the idea of the multiple points cloud collection.TSDF model with 512-pixel in width and height is built in GPU. Given the initial position of the camera relative to the world coordinate system,the coordinate transformation matrix of the camera coordinate system to the world coordinate system can be obtained. The weights are calculated dynamically based on the depths of the current image coordinate points and the model voxel values. The updating will be achieved after several iterations resulting in a stable set of point cloud coordinates. The experimental results show that the recognized target center point relocation accuracy error can be limited within 2.0 mm by introducing the algorithm to the static target grabbing of the manipulator so as to improve the success rate of the robotic grapping target. The performance of the computer will not be affected since the model is built in GPU. The proposed algorithm is more robust than the conventional repairing algorithms,and can be applied for color complex objects and deep images without color images.