In order to achieve accurate picking of citrus picking robots and provide effective guidance for obstacle avoidance path planning during picking

a method three-dimensional reconstruction of citrus trees and mature citrus recognition in natural scenes is proposed. Color image and depth image of the citrus tree are obtained by Kinect v2 and the background is removed by depth threshold segmentation. Point cloud is obtained by color image and depth image registration

and the outlier noise point is removed by the Kd-Tree algorithm in the statistical filter. It is iteratively fused by the ICP algorithm

for the fused point cloud

Poisson’s surface reconstruction algorithm is used for three-dimensional reconstruction of the citrus tree. Combined with the color feature information of mature citrus to segment and identify it under different color spaces. The experimental results show that the precision of citrus tree reconstruction is 94.83%

and the recognition rate of mature citrus is 94.72%.