Research On Tomato Recognition Technology And Picking Path Planning Based On Deep Learning

"Made in China 2025" clearly proposed in agricultural production to enhance information collection technology,intelligent decision-making processing and the ability to accurately conduct operations.The continuous development and renewal of related picking machinery and equipment will promote the development of my country’s agriculture to modernization and intelligence.The main research object of this paper is the automatic identification and picking technology of ripe red tomatoes.Aiming at the problem that the existing technology cannot detect,identify,locate and pick tomato fruits in real time under the complex agricultural environment,a rapid deep learning technology is proposed.Recognize ripe red tomatoes,use image processing algorithms and binocular vision system to accurately locate the recognition target to obtain three-dimensional coordinates,and use matlab to perform simulation experiments of robot kinematics analysis and path avoidance picking planning.The main research contents are as follows:First,analyze in detail the convolution,pooling,fully connected structure and activation functions of the convolutional neural network,compare the target detection algorithms in deep learning,choose YOLOv3 as the detection algorithm,and now collect a large number of different shapes and sizes The pictures of ripe tomatoes were expanded to obtain two thousand valid pictures.Use Label Img to annotate effective images one by one,obtain a data set in voc format,and optimize and adjust parameters,and compare and choose the Keras deep learning framework.Data set training is performed in pycharm through python assembly language to obtain a mature feature detection structure,and finally the algorithm performance verification is performed to ensure that the mature red tomato target can be automatically identified in real time and efficiently.Second,preprocess the real-time image containing the ripe red tomato target,analyze the generation of noise and various noise reduction processing methods,select the combined noise reduction method of principal component analysis and bilateral filtering,and compare RGB,Lab,HSV colors Space,select HSV color space and select the color extraction space range suitable for ripe red tomato fruits,and use color image segmentation method for image binarization.Morphological open and close operations were carried out on the image.And the Canny algorithm is used for edge extraction.Improve the Hough circle transformation,test in the actual environment,obtain the target’s contour and circle center coordinates,and provide basic information for subsequent accurate capture.Third,analyze the principle of the binocular vision structure,choose the Bumblebee2 binocular camera,study the theory and method of dual target calibration and proofreading,use Zhang’s camera calibration method to obtain the camera pose and internal and external parameters of the binocular camera through the dual target toolbox in matlab.The data will be obtained for correction in Open CV.Construct the hand-eye vision system between the camera and the end effector,and convert the pixels,images,cameras and world coordinate systems.The traditional SAD stereo matching algorithm is improved and experiments are carried out to shorten the processing time.The three-dimensional space coordinate value of the fruit is obtained through the binocular camera and the disparity map and the error is analyzed.Finally,the UR10 robot is selected as the research object,the robot establishes the link coordinate system.And the model is established by using the improved D-H coordinate method through the MATLAB robot toolbox.The forward and inverse kinematics of UR10 robot are discussed and analyzed.The trajectory of robot joint space is planned to ensure more stable robot movement.Construct a two-dimensional grid map,improve the ant colony algorithm,optimize the algorithm planning steps,modify its heuristic function and pheromone update optimization and conduct simulation experiments,and then build a threedimensional grid map on this basis to further improve the heuristic function and information The element is updated and optimized,and the motion simulation of the obstacle avoidance picking path of the end effector is performed to simulate the whole process of the robot avoidance picking.