利用超体素中心点连线可见性的图像分割算法

An Image Segmentation Algorithm Using the Visibility of Connection Line Between Supervoxel Center Points

当图像中存在阴影、低对比度边缘和模糊区域时,传统算法利用纹理、颜色等信息难以准确分割物体。以点云图为研究对象,依据盒状、柱状物体在三维空间中的几何形状,提出一种基于超体素中心点连线可见性的图像物体分割算法。该算法首先对点云图进行过分割,得到具有相邻关系的超体素,接着依据其中心点法向量平行准则以及中心点连线可见性准则共同判断相邻超体素是否融合,然后过滤噪声,并利用遮挡关系进一步考察同一物体平面超体素的融合情况。实验结果表明:在OSD-v0.2公开数据集上,文章提出的算法误识别率为6.9%,漏识别率为4.6%,比单一结合法向量的凹凸性判断方法更好,整体误差更小。利用本算法对物体进行分割,更容易计算物体的抓取位姿,有助于机器人执行抓取任务。

If shadow,low contrast edge or blur exists in an image,traditional algorithms face difficulties in segmenting objects accurately by using texture or color,or other information.taking the point cloud as the study subject,according to the geometric shape of box and cylindrical objects in 3D space,an object segmentation algorithm is proposed based on the visibility of connecting line between adjacent supervoxel center points.Over-segmentation is firstly carried out to convert the point cloud into supervoxels with adjacent relationship.Then,adjacent supervoxels are merged together according to the parallel relationship of their normal vectors and the visibility of connecting line between their center points.After filtering the noise,the visibility of connecting line between two planes of an object is taken into account for further merging operation.The experimental results on the OSD-v0.2 open dataset show that false positive rate and false negative rate of our proposed algorithm are 6.9%and 4.6%,respectively,which is better than the concave-and-convex judgment method merging supervoxels by only the normal vector criterion.The overall error of our algorithm is smaller.In addition,it is easier to calculate the grasping pose of an object with our algorithm,which benefits the grasping task.