Recognition And Localization Of Assembly Robot Based On Mask R-CNN

Assembly robot is an important research direction in the field of intelligent manufacturing.With the decrease in population and the introduction of intelligent production,more and more industries and factories have begun to use robotic production.This brings huge development potential to the robotics field,and at the same time increases industry competition pressure.Improving industrial competitiveness is extremely important.Loading machine vision has higher accuracy and adaptive ability,which has great advantages in market competition.However,traditional machine vision has disadvantages.It needs to arrange the targets neatly before identifying and locating the targets,and perform simple contour extraction and feature point matching for each target.If parts are placed in a chaotic state under low light conditions,the recognition efficiency is greatly reduced,which causes certain limitations to the use of machine vision.In order to increase the usefulness of the vision system,the robot can do better in complex environments.This paper applies deep learning to the visual system.The system uses the powerful feature extraction capability of the neural network to obtain part coordinate information.It corrects the picture through camera calibration,and then uses hand-eye calibration to get the coordinates in the robot coordinate system.It can realize the recognition and positioning of the robot.The research contents of this paper are as follows:Firstly,the deep learning algorithms and their advantages and disadvantages are described,and the research status of machine vision and deep learning is analyzed.Finally,the solution of part identification and positioning system was determined.The Mask R-CNN algorithm is selected as the basic algorithm for the problem.The problems of Mask R-CNN algorithm are improved,and GIOU is introduced as the loss function of the improved model.In order to solve the data set problem,transfer learning is used to train the picture data.GPU distributed training method is adopted to reduce the time required for training data.The modified network is first tested,and then Mask R-CNN is trained using the same data set.Because the dataset types of Faster R-CNN are different,the same picture is used to re-create the dataset for training.The results of the three algorithms are compared,and experiments show that the modified algorithm can improve the recognition accuracy and speed.Finally,the mask is extracted and the angle information of the part is obtained using the principle of minimum external moment.The camera was calibrated using Zhang Zhengyous calibratio n method.The system uses Matlab to calculate camera parameters,and uses OpenCV to write programs to implement Zhang Zhengyous calibration method.Compare the results of the two experiments,and save the data obtained by OpenCV as the parameters for image correction.Based on the research of the opponents eye calibration principle,the plane nine-point method principle is used to obtain the transformation relationship between the pixel coordinate system and the robot coordinate system.Finally,the algorithm is implemented using OpenCV and HALCON,and the parameters of the mapping relationship are obtained to realize the transformation of the pixel coordinate system and the robot coordinate system.