%0 Journal Article
%T 基于多层级网络的像素级抓取姿态估计<br>Pixel-Wise Grasp Pose Detection Based on Hierarchical Network
%A 张志刚
%A 程良伦
%J Computer Science and Application
%P 72-82
%@ 2161-881X
%D 2023
%I Hans Publishing
%R 10.12677/CSA.2023.131008
%X <span style=\"font-size:14px;\">为了解决在杂乱场景下从单视角图像中准确估计抓取位姿的问题，本文提出基于多层级特征的像素级端到端抓取检测网络。我们在全卷积神经网络中集成了多层级金字塔池化模块和多分支输出结构形成高精度抓取位姿检测网络，从而有效处理尺寸和位姿各异的未知物体在杂乱场景下的抓取问题。实验表明，我们的方法在Cornell抓取数据集上的理论抓取精度相比现有方法有明显提升；同时，在机械臂实物抓取实验上，我们的方法在多物体杂乱场景中以88.0%的平均抓取成功率实现了100%的抓取完成率。</span><br />
<span style=\"font-size:14px;\">In order to address the problem of estimating grasp pose under cluttered scenes using single-view image, we proposed an end-to-end pixel-wise grasp detection network based on hierarchical features. We integrated a pyramid pooling module and multi-head structure into a fully convolutional neural network to form a high-precision grasp pose detection network, effectively handling the problem of predicting grasps for unknown objects with diverse sizes and poses in clutters. The experiments showed that our proposed method significantly outperforms existing methods in Cornell grasp dataset. Physical experiments on real robotic arms are also conducted, our method achieved average grasp success rate of 88.0% at 100% completion rate in the challenging grasping task in clutter with multiple unknown objects.</span>
%K 机器人抓取，抓取检测，卷积神经网络<br>Robotic Grasping
%K Grasp Detection
%K Convolutional Neural Network
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=60460