基于3D视觉的水泵生产线零部件无序抓取研究

Disorderly grasping parts of pump production line based on 3D vision

为了实现对堆叠零部件的识别、无序抓取及零部件装配状态的检测,依托某水泵生产线智能化改造项目,在现水泵生产线的基础上,建立了一套基于3D机器视觉的堆叠零部件无序抓取系统,从而实现了水泵生产自动化。首先,采用全局特征描述子PPF算法提取了零部件特征;在离线训练阶段,对满足条件的可见点进行了两两组合,计算了点对特征,得到了描述物体全局信息的模型;在线匹配阶段,通过使用基于霍夫变换的投票策略,完成了对零部件的识别;然后,再应用RANSAC算法进行了位姿粗估计,并利用ICP算法对位姿结果进行了微调,得到了目标的最优位姿估计;经过标定的转换矩阵确定了零部件在真实世界坐标系下的位置和姿态,以引导机械手对堆叠摆放的零部件进行精准抓取和放置;最后,将待测状态零部件图像与正确装配零部件图像进行了对比,以判断零部件装配状态,利用水泵生产线改造项目实际搭建了无序抓取系统实验,用以对该系统实施效果进行验证。研究结果表明:以水泵泵体为例,最终选定的点对特征数目为730830,匹配准确率可以达到94.64%,匹配耗时为1.1424 s;总体抓取平均成功率为92.3%,并且识别耗时、位姿粗估计耗时及位姿精匹配耗时均符合实时性要求,抓取系统具备实用性。

In order to realize the identification of stacked parts, the disorderly grasping and the detection of the assembly status of parts, relying on the intelligent transformation project of a pump production line, on the basis of the existing pump production line, a set of disorderly capture system of stacked parts and components based on 3 D machine vision has been established, thus realizing the automation of water pump production. Firstly, the global feature descriptor PPF algorithm was used to extract parts features. Then, in the off-line training stage, the visible points that met the conditions were combined in pairs, and the point pair features were calculated to obtain the model describing the global information of the object;in the online matching stage, parts were identified by voting strategy based on Hough transform. Furthermore, RANSAC algorithm was used for rough estimate of position and pose, and ICP algorithm was used to fine-tune the pose results to obtain the optimal pose estimation. The calibrated transformation matrix determined the position and attitude of the parts in the real-world coordinate system to guide the manipulator to grasp and place the stacked parts accurately. Finally, the images of parts under test state were compared with those of correctly assembled parts to judge the assembly state of parts. Through the water pump production line transformation project, a disordered grabbing system was built for verification. The result of research shows that: taking the pump body as an example, the number of selected point pairs is 730 830, the matching accuracy can reach 94.64%, and the matching time consumption is 1.142 4 s. The average success rate of the overall capture is 92.3%, and the time consumption of recognition, rough estimation of position and pose and precision matching all meet the real-time requirements, which is practical.