| Tomato is one of the most widely planted crops in China’s facility agriculture and its production is an important way for farmers to increase income and get rich.Picking is an important part of tomato production.At present,it is still mainly artificial,with high labor intensity and high cost.The realization of intelligent picking of tomato has become an inevitable trend of development.In this study,the ripening stage tomato was taken as the picking target,aiming at the problems of tomato misrecognition,low positioning accuracy,low picking efficiency and high damage rate in complex environment,the key technologies of the picking robot were studied in three parts: recognition,positioning and picking.The mature tomato recognition model was built,and the target tomato was located based on the depth camera.Finally,the grasp and pick of tomatoes with different diameters were realized by using PWM signal.The main research contents and conclusions are as follows:(1)The tomato recognition model was constructed.Taking YOLOv5 s network as the initial recognition model,the up-sampling operator of the initial model was optimized by CARAFE,which kept the model lightweight and increased the network sensitivity field.EIOU and Quality Focal Loss were used to reconstruct the loss function of the model.The problem of decreasing accuracy caused by unbalanced samples was solved,and HSV was used to segment and detect the detection frame,which improved the recognition accuracy.When the detection category is mature tomato,the proposed YOLOv5s-CLH tomato recognition model has an increase of 1.1percentage points in m AP＿0.5 and 18.1 percentage points in m AP＿0.5:0.95 compared with the YOLOv5 s before improvement.Compared with other detection models such as YOLOv5-lites,YOLOv7,YOLOv8 s and Faster-RCNN,the total m AP value of YOLOv5s-CLH was improved by 11.3,9.3,0.8 and 15.1 percentage points respectively.The recognition accuracy of YOLOv5s-CLH for mature tomatoes in the test set is 97.02%,which was 6.54 percent higher than that before the improvement.The average detection speed of each image was 7.6ms,which can meet the real-time requirements of intelligent picking.(2)The positioning method based on depth camera was adopted.The Real Sense D435 i camera was calibrated with the corners of the chessboard extracted by Open CV,and the internal and distortion parameters of the camera were obtained,and the parameters were used to remove image distortion of the image collected by the camera.The picking test platform was built,and the combination mode of the manipulator and camera was determined to be eye-to-hand,and the positioning formula was deduced to calculate the three-dimensional coordinates of the target to achieve the positioning of the target tomato.(3)The end effector was optimized,and a grasping control method of tomato with different diameters was proposed based on depth information and end PWM signal.According to the size of tomatoes picked,the length,width and camber parameters of the end effector finger part were optimized,and the gripping force changes and clamping injuries of end effector fingers of different sizes were compared through the actual grasping test,and the finger length and width were determined to be 45 mm and 21 mm respectively.At the same time,whether the grabbing effect of the selected end size was optimal was verified by the grasping test.The actual diameter of the tomato was obtained based on the depth camera and positioning formula,and the tomato diameter was converted into PWM signal required for the opening of the end effector and the diameter of the center point according to the corresponding relationship between the PWM signal and the diameter of the tip of the end effector.After moving to the target point,the grasp of tomatoes of different sizes was realized by subtracting the signal difference.The PWM signal difference was determined to be 245.(4)Based on the established tomato picking experimental platform,positioning and picking experiments were conducted in the laboratory,sunlight greenhouse,and multi-span greenhouse,respectively.The positioning error,positioning accuracy,picking speed,and picking success rate were tested.The test results showed that the total average error of positioning after distortion removal was reduced by 6.65 mm,the success rate of positioning was increased by 18%,and the accuracy of positioning was improved.In addition,the highest harvesting success rate of the experimental platform was 91.33%,with an average harvesting speed of 7.95 seconds and a damage rate of approximately 5.33%. |
PDF Full Text Request