Abstract: The problems that the visual system cannot accurately locate parts due to the defects of machined parts have seriously affected the promotion of production automation in small and medium-sized enterprises.To solve this problem,a grasping detection method for abnormal parts was proposed.A Key Point Detection Model(KPDM)based on deep learning was designed to detect the grasping key points of different parts,and then a pose solving module was designed according to the key point position information and hand-eye calibration parameters to calculate the grasping pose of the parts.By combining the architecture of the image segmentation model Deeplab V3+ with the heatmap supervision method,KPDM could capture keypoints from input images.The experimental results showed that the proposed visual grasping system could accurately estimate the position and orientation of parts with different shapes.Taking the electric iron soleplates as examples,the detection success rates for complete and incomplete soleplates in different lighting environments were 97.2% and 92.7% respectively.

Key words: robotic grasp, deep neural network, position and orientation detection, key point detection, heatmap

摘要： 针对加工零件存在残缺导致视觉系统无法准确定位而严重影响中小型企业生产自动化推广的问题,提出一种异状零件抓取检测方法。先基于深度学习设计一个关键点检测模型(KPDM)以检测不同异状零件的抓取关键点,再根据关键点位置信息和手眼标定参数设计一个位姿求解模块以解算出零件的抓取位姿。KPDM结合了图像分割模型Deeplab V3+的架构和热力图监督方式,可以通过输入的零件图像获取抓取关键点热力图。实验结果表明,该抓取方法可以在不同光照环境下准确预测完整零件和残缺零件的抓取位姿,其中两种零件的检测成功率分别为97.2%和92.7%。

关键词: 机器人抓取, 深度神经网络, 位姿检测, 关键点检测, 热力图