基于Mask R-CNN和关键点提取的抓取位姿估计方法

Grasping pose estimation method based on Mask R-CNN and key point extraction

针对工业现场中目标工件抓取位姿任意、在空间中散乱堆叠遮挡的问题,文章提出一种基于掩膜区域卷积神经网络(mask region-based convolutional neural network,Mask R-CNN)和关键点提取的抓取位姿估计方法。对抓取环境中待抓取工件与不可抓取工件采用实例分割,并构建抓取目标表面点云;对目标表面点云进行三维尺度不变特征变换(3D scale invariant feature transform,3D SIFT)关键点提取,对模板点云进行三维角点(3D Harris)关键点提取,以提取的关键点作为采样一致性初始配准(sample consensus initial alignment,SAC-IA)算法的初始值,减少点云配准的计算量,实现目标表面点云与参考模板点云粗配准,利用迭代最近点(iterative closest point,ICP)算法进行精确配准。以三通和弯管作为实验对象,通过与单特征算子提取方法对比分析,表明该方法能够进一步降低点云配准时间,提高配准精度。

For arbitrary grasping pose of target workpieces scattered and occluded in industrial site,a grasping pose estimation method based on mask region-based convolutional neural network(Mask R-CNN)and key point extraction is proposed.The workpiece to be grasped and the workpiece not to be grasped are separated by instance segmentation in the grasping environment,and the point cloud of the grasping target surface is constructed.The key points,extracted from the target surface point cloud by 3D scale invariant feature transform(3D SIFT)and the template point cloud by 3D Harris,are used as the initial value of sample consensus initial alignment(SAC-IA)algorithm to reduce the computational burden of the point cloud registration and achieve the coarse registration of target surface point cloud and reference template point cloud.Then,iterative closest point(ICP)algorithm is used for accurate registration.Taking tee joint and elbow pipe as experimental objects,experiments were conducted by the comparison with single feature operator extraction method,which show that this method can further reduce the point cloud registration time and improve the registration accuracy.