Target Detection Algorithm For Apple Picking Robot In Complex Environment

With the continuous innovation of machine vision technology,many industries has been promoted.And the impact on the agricultural field is also deepening.In the research of fruit and vegetable picking robots,orchard production testing,etc.,the vision system is used to realize target fruit recognition.However,the actual operating environment in the orchard is more complicated that always presents unstructured and uncertain.And affected by the camera position and the growth posture of the fruit tree,the state of the target fruit in the collected images is variable,which seriously restricts the recognition efficiency and accuracy of the target fruit.This study takes the vision system of the apple picking robot as the research object,committed to improving the recognition efficiency and accuracy of the target fruit,and proposed a rapid detection algorithm of the target fruit based on the depth image,innovatively put forward the concept that positioning is before identification.On the other hand,an overlapping target fruit recognition algorithm based on optimized Mask R-CNN is proposed to improve the working ability of the picking robot.The specific research contents are as follows:(1)The target fruit positioning method based on the depth image is to draw an isobath map by analyzing the depth information of the target image collected,obtain gradient field information of the depth image,and project the gradient vector from the three-dimensional space to the two-dimensional space.Rotating the gradient vector in the same direction which will present a vector aggregation area with varying sizes.Because the target fruit is spherical,the center of the regular aggregation is the apple center,while the non-target fruit area is chaotic and divergent.The depth image can quickly locate the center of the target fruit,and is not limited by the color and occlusion of the fruit,which solves the problem of target fruit positioning in complex backgrounds.(2)In RGB space,the graph segmentation algorithm is introduced.By constructing an adaptive threshold and establishing a mapping relationship between the circle center number and k,min,the image is segmented into superpixel regions.The center of the circle obtained from the corresponding depth image then traversed in the superpixel area where the center of the circle is located,which is the target fruit area.The maximum radius of the superpixel area is scanned to used as the circular radius,then,fit the target fruit area by the center of circle found with depth image and the radius found with superpixel area.The new method does not need to design a classifier,and can quickly realize the recognition and positioning of Apple.(3)Aiming at the problem that the target fruit is occluded or overlapped,an overlapping target fruit recognition model based on optimized mask convolutional neural network is constructed.Using deep Residual Network(ResNet)and Densely Connected Convolutional Networks(DenseNet)as a feature extraction network of the overall model to generate feature maps.Input the feature map into the Region Proposal Network(RPN)for end-to-end training to generate the region of interest,and finally Full Convolutional Network(FCN)generates a mask to obtain the area where Apple is located.The deep network structure constructed is continuously optimized through feature extraction and self-learning with features,which greatly improves the accuracy of recognition which target apples are occlusioned or overlapped by branch and leaf.The above research has improved the efficiency of target fruit identification and positioning,the accuracy of overlapping target fruit identification,and promoted recognition effect.The new method can be further extended to the identification and positioning of other spherical fruits,and promote the agricultural production to develop more intelligent.