Research Of Robot Grasping Method Based On Convolutional Neural Network

With the continuous development of social industrial intelligence,the traditional industrial robot has encountered new challenges.Grasping is one of the main capabilities of robots,and the grasping problem of robots has always been a hot topic of research at home and abroad.In recent years,deep learning technology continues to rise.This technology can extract better features,so it is applied in various fields,and scholars at home and abroad also apply it to the autonomous grasping task of the robot.Machine vision technology has developed rapidly in recent years,and the combination of machine vision,deep learning and robot has become a mainstream trend in the development of industrial robots.For this reason,a target detection model based on convolution neural network is designed in this paper,and a robot autonomous grasping system is built.Firstly,thesis studies the main components of convolution neural network,and studies the principle of target detection based on sliding window and the method of realizing full convolution network.This paper also studies the detection principle of Faster R-CNN target detection model and YOLO target detection model,and optimizes the two target detection models.The network structure,anchor scale size,target classification and position regression structure of the Faster R-CNN target detection model are adjusted.The Darknet-19 feature extraction network structure of the YOLO-v2 target detection model,the overlapping confidence of the prediction frame of the NMS algorithm and the loss function are adjusted.Then,the image of the captured object is collected by the Kinect-v2 camera,and the data set is expanded manually.The data obtained are randomly divided into training set and test set according to the proportion of 5:1,and the images in the training set are marked by Labellmg tool.After that,we train and test the two improved target detection models,and analyze the training loss of the two models,as well as the detection accuracy and speed of the two models on the test set before and after the improvement.Finally,the experimental platform is built with Eft robot and Kinect-v2 camera to study the three-dimensional positioning of the object and how to obtain the grasping angle of the object.The Kinect-v2 camera is calibrated by Zhang Zhengyou method,and the hand-eye calibration of the Kinect-v2 camera and the robot is carried out by using the coordinates of several spatial points,and the internal and external parameters of the camera are obtained respectively.After the above research and preparation,the experimental verification is carried out,using three color squares as the target to grab the object,and using two detection algorithms for multiple groups of experiments.The experimental results show that the target detection network learns effective image features,and the grabbing system can basically complete the autonomous crawling task.