Experimental design of robot tracking and grasping based on a 3D point cloud
[Objective]To obtain three-dimensional information about a moving workpiece economically and enable two-dimensional vision to obtain height information about an object,a three-dimensional scanning system based on a line laser is designed.[Methods]The point cloud acquisition function of a moving workpiece on a conveyor belt was developed and realized using a PCL point cloud library.During point cloud acquisition,the laser stripe on the moving object surface has a motion blur effect,which affects the light stripe center extraction.To accurately extract the two-dimensional coordinates of the center from the motion-blurred light stripe image,a threshold method based on the light stripe theory imaging width is proposed.The higher brightness part of the motion-blurred light stripe is retained through threshold processing,which solves the problem of different light stripe imaging widths in the case of motion blur.The experimental results show that the proposed method can quickly extract the two-dimensional coordinates of the fringe center from the motion-blurred light strip image and meet the accuracy requirement by measuring the three-dimensional information of the standard block.At present,the grasping detection algorithm usually needs an accurate object model and a complex matching algorithm or a large number of annotations of the grasping object posture to train the complex detection network to recognize the object posture.Aiming at the problems of low efficiency and cumbersome data annotation of the above methods,a grab location annotation method based on object point cloud information is proposed,which converts the actual grab detection task to the recognition of point clouds.In this method,the feature histogram of the point cloud is used to describe the characteristics of the point cloud of different objects to connect the attitude and grasping position of the actual object with the information about the point cloud.Through the multilayer description of the cloud,we can grasp the information from the cloud to the actual classification.This method does not require an accurate object model or complex matching algorithm,reduces the number of object attitude acquisitions,and avoids cumbersome annotation.[Results]The experimental results show that the constructed multilayer perceptron network can recognize the point cloud information about the object and determine the feasible grasping position from the constructed grasping library.Aiming at the problem that the position motion command of the industrial robot cannot specify the terminal gripper to run at a uniform speed,and the motion command cannot respond in time because of the delay of the upper computer control command,the optimization algorithm based on position control is used to update the tracking path in real time to continuously reduce the distance between the gripper and the object to complete the grasping operation.[Conclusions]The algorithm compensates for the position deviation caused by the sending delay of the upper computer control command,optimizes the robot grasping trajectory,and realizes the nonstop grasping function of the conveyor belt.The experimental results show that the method provided in this paper can complete the nonstop grasping task of real objects.
three-dimensional point cloudlaser fringegrabteaching experiment
万方数据
维普
