Research On Robotic Grasping Dete-Cion Based On Depth Image And Deep Learning

In recent years,population aging and corresponding shortage of labor is becoming increas-ingly serious,as a result,society demonds more service robots.However,The unstructured work-ing environment of service robots brings unknown objects to be detected and grasped.In response to this problem,this paper draws on the great success of deep learning technology in the field of computer vision,and studies the application of deep neural network in robot grasp detection.Based on Baxter robot and Kinect depth camera,a robotic automatic grasping system is established,and two grasp detection algorithms based on deep neural network is deployed in the system.The performance of the two algorithms is tested in experiments.Firstly,the basic framework of the robot automatic grasping system is developed,which is composed of depth camera,robot and grasp detection algorithm.Then,we introduces the prin-ciple of the depth camera module,listes the parameters of the robot and its configurable parallel gripper,and analyzes the 3D and planar representation of the grasp in the deep learning based grasp detection algorithm.Secondly,a two-stage grasp detection algorithm is implemented which first samples the candi-date grasps,then scores the candidate grasp by the deep neural network and selects the best grasp.By designing a gripper width and desktop distance based depth image normalization algorithm and applying it to the Dex-Net 2.0 dataset,the improved Dex-Net 2.0 grasp classification dataset is built.Based on the architecture of GQCNN(Grasp Quality Convolution Neural Network),an improved grasp quality convolutional neural network is designed.By training and verifying on the improved Dex-Net 2.0 grasp classification dataset,the improved grasp quality convolutional neural network yieldes an accuracy of 0.889 and an average precision of 0.859.Thirdly,a one-stage grasp detection algorithm is designed which predicts the grasp feasibility of each region directly from the whole input depth map.The corresponding grasp detection dataset is obtained from the original Dex-Net 2.0 dataset.Based on the idea of one-stage object detection algorithm and anchor,the architecture of the fully convolutional neural network is designed.The angle prediction is divided into multiple regions,which avoids making the strong assumption that a image or a region contains only one grasp.By training and verifying on the grasp detection dataset,the fully convolutional neural network obtains the grasp prediction accuracy as a Dice similarity coefficient of 0.52,the average grasp position prediction error of 0.16 pixels,and the average grasp angle precision error of 5.1 degrees.Finally,the experiment environment and automatic grasping system was established based on ROS(Robot Operating System).In simulation,we collected the test object set from the YCB model set and tested two-stage and one-stage grasp detection algorithm one the test object set in various camera position and orientation settings.As a result,we obtained the average grasp prediction precision of 79%for the two-stage grasp detection algorithm,88%for the one-stage grasp detection algorithm,and the average computation time of 829 ms for the former,4ms for the latter.Using the Baxter robot and the Kinect V1 depth camera,the actual automatic grasping system was established.After calibrating the intrinsic parameters of the depth camera were and the transformation between the robot and the camera,10 objects in daily life were collected to test the well-performing one-stage grasp detection algorithm,and good results were obtained.