Research On Object Recognition And Robotic Grasping Technology Based On RGB-D

The current era is an era of intelligence.Both for traditional industrial robots and service robots that are not yet mature in technology,new challenges are presented.As a very important capability of the robot,robotic grasping and object classification has always been the focus of research at home and abroad.The variety of object types,the variety of object shapes,and the arbitrary placement of object poses impose higher requirements on the robotic grasping and object classification.This paper faces the classification and grasp of common objects in life,using the Kinect v2 camera as the main sensor,focusing on the establishment of the grasp classification model to learn three-dimensional local features of objects and then map grasp parameters of the end of the robot based on local features;for a small number of training samples,this paper studies and analyzes the transfer learning of two classical neural networks to classify items,and sets up an experimental platform for final verification.The work of this paper mainly includes the following aspects:(1)In order to achieve object recognition and classification,this paper study and analyze the transfer learning of two classical neural networks VGG16 and Inception-v3.Analyze the common data set,make some certain processing,and obtain the data sets needed for training and testing.Due to the small number of original data sets,this paper will use the method of transfer learning to train the neural network model;Inception-v3 is selected as the object classification model according to the test results of the two models on the test set.(2)Study four kinds of grasp classification model to obtain grasp parameters of the end of the robot.Four grasp classification models have been established for color information RGB,depth information Depth,four-channel RGB-D information,and two independent input RGB and Depth information.Comparing and analyzing the tests and detections on the datasets that have been marked,the four-channel RGB-D grasp classification model has a success rate of 83.5% in the detection of the top1 rectangular box,which is obviously higher than other models;(3)Rely on Baxter robot and Kinect v2 camera to build experimental platform and verify the feasibility of the model.Using four points in space to complete the hand-eye calibration of the robot,the calibration error is in an acceptable range;Use the 3D point cloud in the top1 rectangle,formulate corresponding rules to map to actual crawl parameters.Then randomly place the four different objects on the desktop to carry out the robot's grasping and objects classification.The experiment proved that the grasp and classification model of this paper is feasible and stable.