Research On Target Location And Estimation Of Grasping Posture Based On Robotic Vision

Robot is the application product of electrical or mechanical automation,known as "the pearl on the crown of the manufacturing industry",is an important symbol to measure a country's innovation ability and industrial competitiveness,has become an important entry point for a new round of global science and technology and industrial revolution.Artificial intelligence is a booster to make the machine more intelligent and humanized.People hope that the robot will behave more intelligent in work just like human beings,which requires artificial intelligence to endow the robot with new life.Deep learning is an important milestone in the development of artificial intelligence and provides a better way for machine learning.Nowadays,robot target grabbing technology is far behind the performance of human beings,which has become an important field and hotspot in robot research.In order to achieve autonomous grasping,the robot first needs to perceive the environment,detect the position of the target object,identify the category of the object,and determine the attitude of grasping the target object,so as to plan the trajectory of the robot and control the robot to grasp.The work of this paper mainly focuses on the problem of target detection and grasping attitude estimation based on robot vision,and combines the theory and technology of deep learning to study and realize the detection,recognition,positioning and grasping attitude estimation of target object based on machine vision.The main work of this paper includes the following three aspects:(1)Binocular camera is used for environment perception to obtain the image and depth information of the object to be captured.In combination with the data set and application requirements in this paper,the neural network model based on regression is used for object detection and recognition,and the migration learning method is adopted to accelerate the training process of the network.The trained neural network is used to detect the position of the object and identify the category of the object,and the image position coordinate and category information of the object to be captured are provided to the robot.(2)The depth information obtained by the binocular camera is converted into point cloud data,and the 3d coordinates of the object relative to the camera coordinate system are obtained by combining the image position coordinates of the object detected.In addition,the coordinate system of the object itself is constructed to provide the pose information of the object for grasping by the robot.(3)In a two-dimensional image of the object,making a more suitable for neural network training of grasping posture,with the method of transfer learning,large data sets for the trained network,based on the design of a kind of more suitable for robot grasping attitude estimation of the network,and in the light of the features of the data set in this paper the design error function,judge whether the grasping posture of the prediction is effective.The experimental results show that the proposed method can improve the grasping precision and robustness of the robot,make the objects captured by the robot more diversified,and the grasping process is more intelligent.