Research On Fruit Recognition And Positioning Algorithm And System Of Citrus Picking Robot

In response to the problem of low picking efficiency of picking robots caused by the low accuracy and slow recognition speed of fruit recognition algorithms in the natural environment,this thesis relies on the key research and development project of Sichuan Provincial Science and Technology Plan to research the recognition and positioning system of citrus fruits in the natural environment,so as to improve the picking efficiency of picking robots and enhance the level of agricultural modernization and automation.The main research content of this thesis is as follows:(1)Citrus dataset production and model selection.The citrus images of different growth stages were taken in a citrus orchard at different light levels,different times and different angles.The citrus data set for model training was obtained by screening,data enhancement and marking the collected images.The model was trained using the Faster-RCNN and YOLOX-S algorithms respectively,and the YOLOX-S model was selected as the base model because its overall efficiency was better than the other models in comparative analysis.(2)The Our-YOLOXS citrus detection model was constructed.To address the problem of leaf false detection caused by the complex background of orchard in natural environment,the ECA attention module is introduced into the feature extraction network to enhance the neural network’s ability to perceive citrus features.To address the problem of picking interference caused by the recognition of fruits and landing fruits on distant fruit trees in natural scenes,the ASFF structure is introduced into the feature pyramid part of the feature fusion network to adaptively learn the feature information of each scale,reduce the recognition rate of non-target citrus,thereby reducing picking interference recognition.To address the problem of missed detection caused by fruit overlap and green leaf occlusion,the loss function was improved in the prediction part,and the calculation method of confidence loss was modified to Varifocal Loss to improve the occlusion problem of citrus detection.The FPS and mAP of Our-YOLOXS model reached 57 and 93.5%,respectively,which were 10%,6.1%,1.7% and 1.4% higher than those of YOLOv3,YOLOv4,YOLOv5s and YOLOX-S models.Finally,the improved model is deployed on the Jetson Nano embedded platform,and the citrus detection experiment is simulated in the laboratory.The correct recognition rate of citrus reaches100 %.(3)Depth ranging based on binocular vision.The characteristics of the mainstream depth ranging scheme are statistically analyzed,and the binocular camera is selected as the depth ranging scheme in this paper.According to Zhang Zhengyou ’s chessboard calibration method,binocular calibration is carried out to obtain the internal and external parameters of the binocular camera.The calibration parameters are used for stereo correction,so that the left image and the right image can be aligned in the same plane.Then,the SGBM algorithm is used to match the left and right images,and the disparity map of the corresponding image is calculated.Finally,the depth distance between each point in the disparity map and the left camera of the binocular camera is calculated by the depth ranging formula.(4)Citrus recognition and positioning system based on Our-YOLOXS and binocular vision.The Our-YOLOXS citrus detection model is combined with the binocular camera depth ranging method to realize the recognition and positioning of citrus fruits.According to the real environment of the picking operation,different distances of 45 cm,60 cm,75 cm,90 cm and 105 cm were set up to verify the accuracy of the recognition and positioning system.The experiments showed that the distance errors of the five groups were 0.1 cm,0.7cm,1 cm,1.9 cm and 2.4 cm,respectively.The error rate of the system was less than 3 %,which could accurately identify and locate citrus fruits.The above research provides practical value for the recognition and positioning system of fruit and vegetable picking robot,and is of great significance for improving the picking efficiency of fruit and vegetable picking robot.