Research On Target Recognition And Path Planning Algorithm Of Citrus Picking Based On Machine Vision

With the development of agricultural automation,the traditional manual picking method has been unable to meet farmers’ demand for picking citrus fruits.There is an urgent need for a fast and intelligent fruit picking equipment to replace manual fruit harvesting.At present,the research of fruit picking equipment in foreign countries has made some achievements,and has been put into practical application,while most of the domestic fruit picking equipment is still in the experimental stage,unable to meet the actual production demand.The research on the algorithm of rapid detection and precise location of picking citrus,as well as picking path planning,can provide the core theoretical support for the picking equipment,which is of great significance for accelerating the research and development of the equipment.In this paper,the target recognition and path planning algorithm of citrus picking are realized based on machine vision technology.The citrus images collected by depth camera are input into the improved Mask R-CNN algorithm for target detection,and then the detection results are transformed into world coordinates and input into the improved ant colony algorithm for citrus picking path planning.The main work is as follows:1.Intel Realsense D435 i depth camera was used to collect RGB images set and depth images set of citrus.Based on the method of median filter,RGB images are preprocessed to highlight the difference between fruits and background.The Label Me tool is used to calibrate the dataset,and a color detection method based on HSV model is proposed to improve the efficiency and accuracy of dataset calibration.The method of image rotation in image processing is applied to enhance the diversity of dataset.2.Through the research of the target detection algorithm of Mask R-CNN,aiming at the phenomenon that the aspect ratio of the extended detection frame is close to(1:1)in the image of citrus,a method of simplifying the RPN network structure is proposed.In this method,the proportion of the preset anchor frame of the original algorithm is simplified from(1:2,1:1,2:1)to(1:1),the prediction accuracy of the algorithm is improved and the convergence speed of the algorithm is accelerated.Finally,test set isapplied to evaluate the improved algorithm model.3.Calculate the barycenter coordinates of citrus mask output by Mask R-CNN as the target pixel position of citrus.By analyzing the working principle of depth camera and the conversion relationship between coordinates,the pixel coordinates of citrus are converted into world coordinates.4.This paper studies the application of genetic algorithm and ant colony algorithm in the traveling salesman problem,and improves the algorithms for the citrus picking path planning,so that they can plan the citrus picking path in the world coordinate system.The optimal parameters of the two algorithms are analyzed and selected respectively.This paper analyzes the path planning effect of the two algorithms under different citrus quantity scale,and selects the ant colony algorithm with better performance as the final citrus picking path planning algorithm.