一种多信息融合的机器人抓取区域检测方法

A Multi-information Fusion Robot Grasping Area Detection Method

重点介绍了面向物体杂乱放置状态下的机器人物体拾取方法。由于物体的多样性和放置时的随意性,很难通过识别、分割、位姿估计等多步骤方法实现杂乱堆放物体的准确抓取。采用一种称为U-net的特殊卷积神经网络、结构融合RGB图像、深度图像信息以及法向量信息,直接预测合适的抓取点,从而无须提前进行物体识别以及位姿估计等操作。最后实验验证了方法的有效性,并比较了多种模态视觉信息对抓取区域预测的效果,发现融合RGB图像、深度图像以及法向量信息可以获得比较好的精度和召回率。

This article focuses on the robot object picking method under the state of disorderly placement.Due to the diversity of objects and the randomness of placement,it is difficult to accurately capture cluttered objects through multi-step methods such as recognition,segmentation,and pose estimation.This paper uses a special convolutional neural network called U-net,structure fusion RGB image,depth image information and normal vector information to directly predict the appropriate grab points,so that there is no need to perform object recognition and pose estimation in advance.operating.Finally,experiments verified the effectiveness of the method,and compared the effects of multiple modal visual information on the capture area prediction.It was found that the fusion of RGB images,depth images and normal vector information can achieve better accuracy and recall.