Abstract: Machine vision is a key technology for robots to identify working objects from complex spatial environments.Kinect depth cameras or laser scanning sensors commonly used in robotic systems are capable of acquiring three-dimensional information about the target,which makes it possible for robots to perform more complex work tasks such as assembly,disassembly,and grasping.However,this also places higher demands on the robot system’s ability to process 3D information such as 3D localization,work object size measurement,and estimation.We analyze the main feature emphasis mechanisms of soft threshold squeeze-and-excitation,channel-wise gated,and attention mechanisms based on PointNet networks,and improve PointNet networks by using soft threshold squeeze-and-excitation,channel-wise gated,and attention networks,respectively,and experimentally validate them on the publicly available ShapeNet dataset from Stanford University.Experimental results show that the improvement of original network by the three emphasis mechanisms improves segmentation accuracy(mean intersection and merge ratio) of 3D point clouds by 0.24%,0.68%,and 0.93%,respectively,in comparison with original PointNet network.The improved method lays foundation for the subsequent solution of accurate estimation for the size of working objects in tasks such as assembly,disassembly and grasping by robots.

Key words: Machine vision, 3D point cloud, Squeeze-and-excitation, Channel-wise gated, Attention module