Research On Detection Method Of Citrus In Natural Environment

As an important fruit producer in the world,China' s citrus yield has reached 38.16 million tons by 2017.However,due to the complex fruit picking environment,fruit picking is still dominated by manual picking by now.With the rapid development of urbanization in China,the sharp decline of agricultural employees and the rise of labor costs make the fruit industry face the problem of labor shortage,which is not conducive to the rapid development of the fruit planting industry.Realize automatic fruit picking is of great significance for solving the shortage of labor force in fruit industry and improving the market competitiveness of fruit.Citrus recognition in natural environment is an important supporting technology for automatic picking of citrus.Most of the existing citrus target recognition algorithms design the target recognition algorithm based on the color,texture,shape and other visual features of the target.Natural environment is complex and unstructured,detection target will be affected by illumination change,uneven brightness,similar color of fruit and background,shade shadow factors.So that its appearance features have great changes,along with the change of environment lead to difficulty in extracting the environmental interference features,the goal of complete feature set in the case of multiple interference factors appear at the same time,the fruits of the existing algorithm target recognition efficiency are poor.Deep convolutional neural network has been proved to be an effective image target feature extraction network,which can extract low-level features from a large number of training data through convolution structure and obtain high-level features through pooling structure.Semantic features obtained by feature learning method have good adaptability to complex environment,which provides a good foundation for improving the accuracy of target detection in complex environment.In order to guarantee the detection system still has high detection accuracy under illumination change,shadow,color changes,the branches shade and fruit overlapping,this paper designed based on the depth of the convolution network under the condition of natural harvest citrus target recognition algorithm,within the scope of the vision to identify citrus;At the same time,in order to ensure that the picking end can be quickly aligned with the target,this paper designed a target scale adaptive KCF tracking algorithm to guide the robot picking end equipment to move towards the orange target.Through the collaboration of the two methods above,the goal detection of the orange picking robot in the natural environment was realized.After the test,compared with classical target recognition algorithm--Deformable part model(DPM),the paper designed for target detection based on the depth of the convolution network system has good ability of target detection in a complex environment,the target scale adaptive KCF tracking algorithm in mobile is good at the end of the picking robot target tracking.In this paper,a target data set conforming to VOC2007 format is constructed based on the citrus target samples collected in the actual picking site,and the algorithm designed in this paper is tested on this data set and good recognition results are obtained.The method proposed in this paper has a good adaptability to the natural picking environment and provides a good method support for the design of practical citrus automatic picking robot.