Research On Citrus Picking Robot Branch Detection Method

Citrus is one of China's important fruit products,and its world trade volume also ranks among the top three.At present,citrus picking is still mainly manual,especially in hilly areas.Because of its large slope and mostly living in high mountains,injuries are more likely to occur.Therefore,the research team designed a citrus picking robot suitable for this situation.The first-generation robot developed can achieve obstacle-free fruit picking.For complex citrus orchards,fruits are often blocked by branches and obstacles.Therefore,how to obtain information about the stems of fruit branches to guide the robotic arm to avoid obstacles has important research significance.In this paper,the citrus fruit twigs in the natural environment are taken as the research object,and the value stem recognition and localization research based on convolutional neural network and Kinect V2 camera is performed.The research content is as follows:(1)Choice of target recognition algorithm.The structural modules of convolutional neural network are introduced,and the main target detection algorithms based on neural network are analyzed,including YOLO algorithm and Mask RCNN algorithm.For each algorithm recognition effect and the purpose of this paper,the Mask RCNN algorithm is determined.As the base model.The main reason for choosing this algorithm is the mask branch added by the algorithm.The obtained mask can accurately segment the target.Based on this,in this paper,the minimum circumscribed rectangle processing is performed on the mask.The obtained circumscribed rectangle frame can more easily obtain the deflection information of the branches.At the same time,the results of the circumscribed rectangle processing also paved the way for subsequent research.(2)Production of training data set.The previous research pointed out that the mask needs to be processed.How to make the processed branches have deflection information requires a better data set production criterion.In the collection of the data set,in order to ensure that the data set and the robot are in the same environment,the shooting distance between the camera and the target is determined.In the production of the data set,in order to unify the scale of the image mark,the minimum scale range of the image mark is determined.At the same time,two types of markers are designed using discrete markers.The shape of the unbranched area is similar to a rectangle,and the shape of the branched area is similar to a trapezoid.(3)Two-dimensional reconstruction of fruit tree trunks.Based on the analysis of the circumscribed rectangular frame,a number of constraints were determined,namely distance constraints,angle constraints,and intersection constraints.Through these constraints,this paper also proposes a multi-parameter constraint partition algorithm,which can complete the orderly connection of discrete borders.Four-degree polynomial fitting is performed on the divided ordered points,and finally the two-dimensional branch reconstruction is realized.The branch fitting error is 11.47%,and the average accuracy of overall branch reconstruction is 88.64%.(4)Three-dimensional positioning of branches.Calibrate the Kinect V2 camera and obtain the internal and external parameters of the camera.At the same time,the space depth constraints are determined,which can better realize the space division of the frame.Through the camera spatial position conversion relationship,the construction of the three-dimensional model of the branch is finally achieved.Through the comparison experiment between the measured value and the calculated value,the deviation of the depth value of the constructed branch information in the real space is less than 1.8 mm,and the average relative error of the branch diameter is 5.6%.Outdoors,the citrus picking robot picks fruits with branch obstacles.The picking success rate is 84.48%,and the obstacle collision ratio is 13.46%.