The Approach Of Orange Recognition And Location Using Lightweight Neural Network And Binocular Vision

Orange is the largest category of fruit in the world.It is the third largest trade agricultural products in the world and it is an important pillar of Chinese fruit industry.Traditional orange orchard picking mainly relies on manual work,which is inefficient,labor intensive and labor costly.Therefore,research on picking robot is an important development direction to liberate rural labor force.In the field of picking robot,orange fruit recognition is the key technology to realize the precision management of fruit trees and the accurate prediction of orchard yield.Orange fruit location is the necessary prerequisite for automatic picking of the manipulator.Moreover,the accuracy of fruit recognition and location directly determines the success rate of picking of the manipulator.In order to achieve the fast and accurate recognition and location of fruit by fruit picking robot,this paper takes oranges as the research object and studies the target recognition and location method using lightweight neural network and binocular vision.The main research work and conclusions are as follows:(1)Illustrated the common algorithms in the field of target detection.In order to realize rapid and accurate recognition of orange’s fruits in natural environment,a method of orange fruit recognition was proposed based on improved YOLOv3-LITE lightweight neural network.In this method,the regression loss function of GIoU border was introduced to improve the accuracy of fruit recognition regression box.In order to facilitate migration to mobile terminals,a lightweight YOLOv3-LITE network model was put forward.MobileNet-v2 was used as the backbone network of the model,and a pre-training method combining mixed training and transfer learning was used to improve the generalization ability of the model.(2)To realize the three-dimensional spatial location of oranges,a spatial location method based on binocular stereo vision was proposed.In this method,the internal parameters,external parameters and relationship coefficient matrix of left and right camera were obtained by calibrating binocular camera.Making use of the SGBM stereo matching method in OpenCV computer vision library to extract features as well as the corresponding feature points of pixels in left and right eye images were matched.(3)The recognition method and location method of orange fruit were tested and verified respectively.In the orange fruit recognition experiment,by comparing the recognition effect of the model with Faster R-CNN and SSD model under the test samples with different degrees of occlusion,the difference of each model was evaluated with F1 value and AP value,the experimental results showed that the recognition effect of YOLOv3-lite lightweight network model was significantly improved.The F1 value and AP value were 93.69% and 91.13% respectively,and the Average IoU was 87.32%.On the GPU,the detection speed of orange target can reach 246 frames/s,the inferential speed of the single 416*416 picture is 16.9 ms,and the model occupies 28 MB of memory.Using binocular stereo vision technology to conduct three-dimensional spatial location experiment for the center point of orange fruit,the error range of X coordinate using 3D spatial positioning method is-5.3～4.6 mm by comparing high-precision laser rangefinder.The error range of Y coordinate is-4.9～4.5 mm.The error range of Z coordinates is-7.1～11.2mm,which meets the 3D positioning error requirements of the picking robot.