Research On Target Recognition And Localization Technology Of Citrus Harvesting Robot Based On Binocular Vision

Citrus is one kind of the widely cultivated fruit in our country.During citrus production,harvesting is labor consuming,accounting for 33 percent to 50 percent of the workload.Therefore,the demand for labor is huge.In order to realize mechanization of harvesting,intelligent harvesting robot is introduced into the citrus production process.Intelligent harvesting can liberate the limited labor resources and improve product ivity,which not only has important practical significance,but also is the inevitable choice of agricultural modernization.The identification and location of the fruits is the primary task and design difficulty of harvesting robot.Therefore,it is of great significance to study target recognition and location of citrus harvesting robot.Based on binocular vision technique,target recognition and location for citrus harvesting robot was studied and valid recognition and location in three-dimensional space of citrus was realized,which provides necessary information for robot.The main achievements of our research are listed as follows:(1)A binocular stereo vision system is established and binocular cameras are calibrated,through which internal and exter nal parameters are defined.The rotation matrix and transformation of spatial position relationship of two cameras are described after calibration of left and right camera stereo is accomplished.(2)Segmentation algorithms of citrus images are studied.Co mpared with three different segmentation algorithms,K-means clustering algorithm and threshold segmentation of HSV based on color space are adopted.Image preprocessing method is designed for further process of the segmented image.(3)Recognition methods of identifying single citrus and overlapped citrus are studied respectively.A method based on based on convex hull and distance transformation theory for localization of overlapped citrus is proposed to determine the center of picking point and reappear overlapped targets of citrus.The experiments demonstrate that the average recognition error is 2.03% for single citrus,the average error of center picking point in the simulation experiments is 6.51% and the average error is 1.58% in the actual experiments.The average error is 13.78% in the experiments of reappearing overlapped citrus.The algorithm could accurately recognize citrus.(4)Three-dimensional localization method for citrus is studied.SURF algorithm is used to extract feature points and RANSAC algorithm and epipolar constraint are applied to remove wrong matching points.A matching method of picking center points based on similarity of image and epipolar constraint is proposed to realize calculation of three-dimensional coordinates of center picking points of citrus targets,and experimental results show that average positioning error of this method is 1.842 mm,which meets the positioning requirements of picking robot.