Research On Identification And Positioning Technology Of Passion Fruit Based On Embedded Platform

With the deepening of the rural revitalization policy,our country in agricultural crop production is rising year by year,especially fruit yield,however the current our country most the picking fruit in the supply chain link is still mainly rely on artificial,not only inefficient,and with the rising of labor costs,is bound to hinder the further development of fruit industry,unable to meet the market’s supply and demand.In order to overcome the above limitations,the use of fruit picking robot instead of manual picking has gradually become a research focus,and will become the development trend of orchard agricultural management modernization in the future.Among them,the identification and positioning system is the key of robot picking,and it is difficult to judge the complex target object,calculate the target depth value,and integrate the binocular ranging model with the target detection model.At the same time,the embedded landing should be taken into account.Its breakthrough in technology will effectively improve the performance of fruit identification and positioning,so as to improve picking accuracy,improve efficiency,reduce costs and drive the further development of the fruit industry.In this paper,the fruit target detection model based on deep learning network and target location algorithm based on binocular vision were studied and improved to solve the problems of low recognition of complex target of passion fruit,low calculation accuracy of target depth value and combination of ranging model and target detection model in natural environment.The experimental results showed that the horizontal recognition accuracy of target passion fruit was the highest,with an average relative error of 10.32%.The average relative error is 12.85%when it is located directly below.The results lay a foundation for the realization of passion fruit picking robot.The main research work of this paper is divided into the following three parts:(1)Fruit target detection.A passion fruit detection model based on YOLOv3 was proposed.According to the characteristics of passion fruit target size,the improved K-means++ algorithm based on cross ratio as distance measure was used to retrieve anchor selection boxes matching the target fruit.The accuracy of target selection and the convergence rate of model are improved.Secondly,in the output network,soft-NMS algorithm used to screen the target prediction box is used to improve the suppression parameters of its Gaussian function in the form of linear function,so as to improve the adaptability and detection ability of the model in different intensive scenarios.Finally,the enhanced YOLOv3 model was used for experimental comparison on the pre-processed passion fruit data set.The experimental results showed that: The m AP and F1 values of the enhanced YOLOv3 target detection algorithm reach94.62% and 94.34% respectively,which are 4.58% and 3.68% higher than the original YOLOv3 algorithm.The average detection speed is25.45 frames /s.(2)The method and realization of passion fruit location based on binocular vision.Firstly,the basic principle and process of binocular vision ranging are analyzed.Secondly,the lens parameters of binocular camera are obtained by Using Zhang Zhengyou calibration method in MATLAB platform.The reprojection error is 0.07,which solves the problem of binocular vision ranging accuracy becoming worse caused by camera distortion.Furthermore,the basic implementation process of the parallax matching algorithm is analyzed,and the SGBM stereo matching algorithm is studied and implemented.The algorithm has the advantages of good stereo matching effect and moderate running speed,which is suitable for the needs of embedded systems.Finally,it was verified by Python language on Py Charm platform,and the experimental results showed that the above algorithm fully met the functional requirements of high-precision target localization of passion fruit,laying a foundation for the implementation of the algorithm in embedded platform.(3)Realization of software and hardware system of fruit recognition and location based on embedded development platform.First,the basic installation of NVIDIA Jetson TX2 platform operating system and environment is completed,and the deployment of Python development environment is completed based on Deep Stream framework,and the decoding and acceleration of video stream is realized.Secondly,the YOLOv3 target detection algorithm is transplanted into the deployed environment,and the training of YOLOv3 is completed on the embedded system.Finally,based on SGBM stereo matching algorithm,passion fruit target recognition and location was realized.Experimental verification shows that the system can basically meet the requirements of real-time fruit recognition and location.Within the recognition range of 90 cm,the average recognition rate of passion fruit is 97%,the average frame rate is20 fps,and the average relative error of distance measurement is 11.30%.