Abstract: Aiming at the problem that the target object was partially occluded during the capture process,a deep convolutional neural network based on multi-scale feature fusion was designed to extract the projection feature points of the three-dimensional target,and the different Perspective-n-Point (PnP)algorithm was used.The network uses synthetic data automatically generated by the computer for training.After verification,the network trained with synthetic data could also work effectively in real scenes.Finally,a UR5 robotic arm grasping platform was built based on Robot Operating System (ROS).The trained model was deployed on the platform for grasping experiments to verify the actual application effect of the pose estimation algorithm.Experiments showed that the proposed method using only RGB image information could estimate the pose of the target object,and capture the object with unknown pose in the actual scene.

Key words: deep learning, pose estimation, perspective-n-point algorithm, grab of manipulator

摘要： 针对抓取过程中目标物体部分被遮挡的问题,设计了一种基于多尺度特征融合的深度卷积神经网络提取3D目标的投影特征点,并根据不同投影特征点数采用不同的多点透视成像算法。网络使用计算机自动生成的合成数据进行训练,经过验证,使用合成数据训练的网络也能在真实场景中有效工作。最后,搭建了一个基于机器人操作系统的UR5机械臂抓取平台,将训练好的模型部署到该平台上进行抓取实验,结果表明所提方法能够估计出目标物体的位姿,并在实际场景中抓取位姿未知的物体。

关键词: 深度学习, 位姿估计, PnP算法, 机械臂抓取