Object Visual Detection And Localization For Manipulator Grasping

The visual assisted manipulator autonomous grasping technology is one of the research hotspots in the field of automatic industry.Currently,the target detection and localization tesk of commonly used visual assisted grasping system is achieved via twodimensional images.However,due to the lack of depth information in 2D images,this kind of grasping systems are usually applied to the structured scene with limited target categories and relatively fixed poses.Compared with 2D images,the 3D images contain extra depth information,are able to describe the shape feature and spatial position of the target more efficiently.Therefore,the 3D visual grasping system is more adaptable to the unstructured scenarios with diverse target species and variable poses,which can significantly improve the production efficiency.This paper focuses on 3D visual detcetion and localization tesk,and studied various related technologies in the robotic autonomous grasping system.The main work and research contents are as follows:(1)Collect the 3D image(in the form of point cloud)of the scene and extract the interest region which contains the target object;Firstly,by taking Microsoft Kinect as example,the process and principle of camera internal parameter calibration are summarized.Then,several pre-processing algorithms are applied on the original point cloud to optimize the data quality and reduces the irrelevant data such as pass-through filtering and voxel sampling,et al,which lays a good foundation for subsequent processes.(2)An improved 6D pose estimation algorithm based on point pair feature voting frame is proposed.Firstly,an novel down-sampling strategy is proposed in the preprocessing stage to avoid introducing a large number of similar surface points while maintaining the surface shape character.Secondly,in the online identification stage,series improvement measures,such as random voting,staged voting and blocking scene coincidence points,are put forward to reduce the calculation amount and speed up the identification process.Finally,the recognition and registration experiments are carried out on public data set and actual scene respectively to verify the efficiency and accuracy of the proposed algorithm.(3)An physical visual grasping system based on Baxter robot is constructed.The hardware platform of the visual grasping system is constructed with Baxter manipulator,Microsoft Kinect depth camera and a PC with Ubuntu OS.Moreover,a corresponding software system which contains multiple functions including object recognition,6D pose estimation,hand-eye calibration,grasping pose detection and trajectory planning,is designed based on ROS,Move It and PCL.