Research On Target Recognition And Capture Based On Multi-angle Splicing Point Cloud

Nowadays,with the rapid development of science and technology,the demand for labor is also increasing.Robots are rapidly occupying the fields of industry and service industries.The technology of sensor's ability to perceive space continues to mature.The research direction of perception through RGB-D information is becoming more popular.To achieve the identification and location of the target applied in various fields.Traditional industrial robots only need to complete simple grasping and placing tasks,while service robots not only need to obtain the target pose,but also need to complete the recognition of the target.In the field of target recognition,the development of two-dimensional image recognition technology is relatively complete,but RGB-D information can describe the object more completely,and the three-dimensional reconstruction of the target can be performed through observations from multiple angles to obtain complete object description information and improve the object The recognition rate.Therefore,this paper uses a depth camera as a vision sensor,which is mounted on the end of the robotic arm to complete multi-angle observation of the target,obtain complete target information,complete target recognition and pose estimation of the object,and control the robotic arm to reach the grasping pose.The content of specific research is as follows:Firstly,the kinematics model of the manipulator is studied.The kinematics analysis of the manipulator is completed.On the basis of studying the imaging principle and depth measurement principle of the depth camera,the calibration of the depth camera was completed,and the depth camera was mounted on the end of the manipulator for hand-eye calibration,so as to determine the relative position of the depth camera and the end of the manipulator in the simulation environment.Secondly,this paper studies the method of obtaining complete target point cloud information.First of all,control the robot arm to move to the initial pose to obtain the initial scene of point cloud,using the plane fitting point cloud segmentation method to identify the target point cloud was isolated,get its external contours and centroid coordinates,based on the centroid coordinates to solve multiple observation angle of the mechanical arm position at the end,control arm movement in turn to each observation position,get the position of scene of point cloud,the initial scene of point cloud as target point cloud,the multi-angle point clouds are spliced by color point cloud registration method,and then the plane in the whole scene is filtered out to get the complete point cloud information of the target.This paper studies the deep learning network structure that can be directly used to process point cloud data.Using its classification network framework,the complete point cloud of the target is input into the network to obtain the target recognition result,and the target pose and the end of the robotic arm are calculated according to the recognition result.According to the minimum bounding box of the target,the robot arm is controlled to complete the obstacle avoidance planning movement to the position to be grasped and the target grasping position.Finally,this paper designs the complete experimental process and builds a complete experimental system.The robot operating system is used as the control center of the virtual motion of the manipulator,and Rviz is used as the visualization tool for the computer terminal manipulator motion display,and the observation is verified in the Gazebo simulation environment.Angle solving,point cloud stitching,target recognition,and pose estimation algorithms have completed the target recognition and capture experiment.