Research On The Pose Detection For Fruit Sorting Parallel Robot Based On Binocular Vision

Compared with manual sorting fruit,sorting fruit through robot has the advantages of high sorting efficiency,low labor intensity and good economic benefits.The parallel mechanism adopts closed chain structure and has the advantages of stable structure,high stiffness,high precision and good dynamic performance.In this basis,a fruit sorting parallel robot is developed by our research group.When the fruit sorting parallel robot sorts the fruit,the pose detection for end effector is beneficial to master the motion state of end effector of the robot and realize real-time closed-loop control,so as to improve the control performance of the parallel robot.The detection method based on the binocular vision has the advantages of non-contact,strong applicability and high cost-effective.Therefore,the pose detection method based on the binocular vision is suitable for the fruit sorting parallel robot.At present,there are still some difficulties in the study of pose detection for the parallel robot based on the binocular vision,such as the decrease of pose detection accuracy caused by illumination,noise interference and other factors,the error pose detection in visual blindness caused by the calibration plate fixed on the end effector being easily obscured by the rod of parallel robot mechanism at the boundary of workspace.For the issues above,a pose detection method based on the improved PROSAC algorithm is proposed to overcome the decrease of pose detection accuracy caused by illumination,noise interference and other factors.And a pose detection method in visual blindness is proposed by combining the forward kinematics and the hybrid optimization RBF neural network for error compensation to solve the problem of error detection in the visual blindness.The main work are as follows:(1)A pose detection method based on the improved PROSAC algorithm is proposed.The influence of factors such as illumination and noise interference can cause the decrease of pose detection accuracy for the parallel robot based on the binocular vision.To solve the above problem,the PROSAC algorithm is improved via selecting feature points separately and pre-testing candidate model.The improved PROSAC algorithm can overcome the low estimation accuracy with respect to the model parameters and time-consuming on testing the error candidate model.Then,the improved PROSAC algorithm is used to purify the stereo matching results to improve the accuracy when guaranteeing the real-time of the stereo matching,so as to the improve the accuracy of pose detection for the fruit sorting parallel robot based on the binocular vision.(2)A pose detection method in visual blindness is proposed by combining the forward kinematics and the hybrid optimization RBF neural network for error compensation.The error detection problem in visual blindness can be caused by the calibration plate fixed on the end effector being easily obscured by the rod of the parallel robot mechanism at the boundary of the workspace.To solve the problem,the forward kinematics is adopted to achieve the theoretical estimation of pose parameters and a hybrid optimization RBF neural network is designed to compensate the error of the the theoretical estimation of pose parameters to improve the accuracy of pose detection in visual blindness.(3)The experimental platform of pose detection for the fruit sorting parallel robot based on the binocular vision is contructed,and the experiment is carried out based on the platform.The hardware platform is designed to collect and transmit the images of end effector.The software development of pose detection based on the binocular vision for the fruit sorting parallel robot is realized via combining Visual Studio Software and Open Computer Vision Library.On this basis,the pose detection experiment for the fruit sorting parallel robot is completed.Experimental results show that,compared with the pose detection method based on the unimproved PROSAC algorithm,when the pose detection method based on the improved PROSAC algorithm is applied,the mean absolute value of error for pose component x,pose component y,pose component z and pose component ? are reduced by 65.9%,68%,66.7% and 58.7%,respectively;the standard deviation of error for pose component x,pose component y,pose component z and pose component ? are reduced by 63.2%,61.9%,63.6% and 56.5%,respectively.Compared with the uncompensated forward kinematics,when the pose detection method combining the forward kinematics and the hybrid optimization RBF neural network for error compensation is applied,the mean absolute value of error for pose component x,pose component y,pose component z and pose component ? are reduced by 54.4%,67.7%,54.7% and 52.9%,respectively;the standard deviation of error for pose component x,pose component y,pose component z and pose component ? are reduced by 52.9%,62.8%,51.9% and 58.8%,respectively.The results verify the effectiveness and superiority of the proposed methods.