Research On Grasping Control Based On Robot Vision Recognition

As the new round of global technological revolution and industrial change continues to deepen,artificial intelligence,image processing and other technologies have been rapidly applied in the field of industrial robot grasping.However,there are still some problems such as poor localization accuracy,inaccurate classification and recognition for the target grasping objects.Aiming at these problems,this paper designs a grasping detection algorithm based on YOLOv3 network,further proposes a grasping location recognition network based on multi-task convolution according to the actual application requirements of industrial robots,which can simultaneously conduct joint detection of grasping location tasks and classification recognition tasks.Finally,a grasping experimental platform based on robot vision recognition is built and the accuracy,instantaneity and feasibility of the network designed in this paper are verified through the experiments.(1)An ED-YOLO network based on edge extraction and dilated convolutions is designed in this subject.The method uses the stable YOLOv3 network as the backbone,carries out multi-channel feature fusion based on edge extraction,and adds Edge Feature module to improve the network’s detection accuracy and velocity.Based on the dilated convolutions to improve adaptively spatial feature fusion,a Dilation-ASFF network is formed to replace part of the original network,which improves the network’s ability to sense multi-scale information and the detection accuracy.CIoU is used to modify the grasping frame regression loss function to improve the convergence ability and accuracy of the network.The improved algorithm achieves an AP of 96.79 on the Cornell grasping detection data set and a detection speed of 0.063 s per sample,which is 2.98% higher than the original YOLOv3 network AP.(2)Aiming at the actual requirements of grasping task in practical engineering,a MT-ED-YOLO network for grasping localization and recognition based on multi-task convolution is proposed in this subject,which can simultaneously complete grasping frame detection,grasping object recognition and classification detection in one model.Based on ED-YOLO network,this network maps the multi-scale shared convolutional feature maps extracted by Dark NET-53 feature extraction network to two branches,EDYOLO and Multi-Task,respectively carry out grasping target localization and item classification tasks to achieve multi-task output of the same network.Meanwhile,in order to better adapt to the multi-task network characteristics,an alternating training strategy combining full parameter update and exclusive parameter update is used to enable the network to converge quickly and stably for the loss of each task during training.Finally,the detection accuracy of grasping frame reached 95.85% in Cornell grasping detection data set,and the item classification accuracy reached 93.34% in self-built item classification data set,which verified the effectiveness of the multi-task network structure and the training strategy.(3)A set of experimental grasping platform based on robot binocular vision is built in this subject.In this experiment platform,binocular camera was used to obtain image information,Zhang’s calibration method was used to calibrate the parameters of the camera.Matlab was used to complete the modeling and simulation of the 4-DOF manipulator,which verified the feasibility of the forward and inverse solutions of the manipulator.A 4-DOF grasping manipulator was used to carry out several object grasping experiments,and the effect of multi-task grasping localization recognition network MTED-YOLO in practical application was tested.The experiment’s final localization success rate is more than 90% and the grasping success rate is more than 80%,which verifies the effectiveness of the algorithm proposed in this paper.