| Citrus are one of the large-scale agricultural products grown in China and an important pillar of agriculture and economic development.However,at present,the picking of citrus fruits mainly depends on People,and the workers are labor-intensive,which leads to high production costs and low productivity.Therefore,the development of high-level picking robots instead of manual operations is very important to liberate people from complex agricultural production.It is of great signification to develop high-level picking robots to take place of manual operations,which can liberate people from complex agricultural production.The visual identification system can accurately identify the position of citrus fruits and acquire the information of obstacles(acquisition of citrus fruits and surrounding obstructions),which is the key to ensuring the successful picking of robots.This article focuses on the following aspects of the target recognition and location methods for citrus picking robots in the natural environment:(1)According to RGB,HSV,Lab and other color space characteristics,selecting a reasonable color space.For the original image captured by the camera,there are problems such as noise interference.By analyzing the noise characteristics,image enhancement and median filtering are used to preprocess the image,which reduces noise interference and increases image contrast.(2)Focusing on the problem of muti-objective classification and identification of citrus in natural environment,a deep convolutional neural network object detection algorithm was used to identify and locate citrus and the type of obstacles.The color characteristics and texture features of citrus in natural environment were analyzed.The SVM segmentation method based on regional characteristics was used to obtain citrus fruit region.The least squares method was used to ellipse the segmented region to restore the real citrus fruit,and the center point of picking was obtained.The experimental results show that the accuracy rate of citrus recognition is 86%,accuracy of Branch classification is 59.5%,while its error rate is about 7.2%.(3)The three-dimensional positioning method of citrus fruit was studied.The binocular camera was calibrated to obtain the internal parameters and external parameters of the left and right cameras.The left and right cameras were stereo-calibrated to obtain thespatial position relationship matrix of the left and right cameras.Using a deep convolutional neural network object detection algorithm to detect the citrus regions from the left and right images,and calculating the average region,as a matching primitive;the MAD with an accuracy rate of 98.5% was selected as a matching metric function by experiment;the left and right images were obtained by stereo correction.The pixels of the line are kept in the same horizontal position.Horizontal line polar lines are added,and the size and type of the area detected by the left and right images are used as constraints to shorten the matching time and improve the matching accuracy.The experimental results show that the average error rates of the x,y,and z axes are respectively 4.32%,3.86%,and2.85%,and the average error sizes are 2.88 mm,1.946 mm,and 1.55 mm,respectively.(4)Using a fixed hand-eye calibration method,the conversion relationship between the camera coordinate system and the robot coordinate system is obtained,and the three-dimensional space coordinates obtained by the camera are converted to the three-dimensional coordinates in the robot coordinate system to provide the picking target point for the picking robot.A citrus picking experiment environment was set up to verify the accuracy and effectiveness of the citrus recognition,obstacle classification,three-dimensional spatial location,as well as coordinate transformation of the picking robot vision system.The experimental results showed that the vision system of the citrus picking robot in this research group is 80% in the success rate of picking citrus.The success rate of obstacle avoidance is 60%. |
PDF Full Text Request