Research Of Vision System Using In Peony Fruit Harvest Robot

The oil peony fruit’s picking operation in our country is done manually which leads the low degree in mechanization and intellectualization.In order to reduce the labor intensity,improve the picking efficiency and then preferably develop the oil peony industry economy.This paper focus on the study of oil peony fruit harvest robot’s vision system,thorough analysis its hardware structure and software algorithm.To solve the critical technical problem of visual system,the study of fruit’s recognition,localization and maturity classification combined with the fruit’ s characteristic are proposed.The research provides the essential theories and technological base for the development of harvest robot.1.The establishment of a vision system:Combined with the oil peony fruit’s growth environment characteristics,the multi-source visual system is established which involves PMD Camcuber 3.0 depth camera and Logitech C270 color camera.It could obtain the color information,amplitude and depth information.And the design remedies the traditional vision systems with their weakness of sensitive to light,position error and poor real-time.The system provides a basis for fruit’s recognition,localization and maturity classification.2.Feature points obtaining and matching:To solve the problem of cameras’ matching,the color camera’s grayscale image and depth camera’s magnitude image are matched.The detection algorithm of Harris and SURF feature points description are compared and analyzed.And according to the results,this paper finally selects the SURF algorithm to detect the feature points.Then,the ratio of the closest neighbor and second closest neighbor is used in the features matching;the Hessian matrix trace accelerates the matching during the process.Finally,the RANSAC algorithm is applied to remove false matching points and then output the matching images.3.Fruit’s recognition and localization:On the basis of images matching,NDLT algorithm is used to correspond these images which have different resolution ratio.It lays the lay a foundation for fruit’s recognition,localization and classification.A partition method for G components of depth image is proposed.Combined with the characteristics of roundness and normalized central moments,a recognition method based on Multilayer Perceptron(MLP)is also proposed.For fruit localization,the centroids are calculated and then compared with actual centroids position.The result shows the algorithm is accurate.4.Maturity classification:an auto-classification method based on MLP is proposed.For tree mature stages peony tree fruits(Green mature stage,yellow mature stage,fully mature stage),it constructs the MLP model and sets the color characteristic and texture characteristic as input.R-G,R+G,H as color features are used to classify the green mature stage fruits.Then,texture features in range highlight is extracted and area threshold function is set to classify the yellow and fully mature stage fruit.Experimental results show that the MLP model for oil tree peony fruits could successfully classify three stages fruits and the accuracy reaches 93.45%.