Study On The Key Technologies Of Intelligent Pruning For Dwarf And Dormant Jujube Tree

In view of high labor intensity and cost for manual pruning,and wrong and missing pruning caused by mechanical pruning for dwarf and dormant jujube trees,study on automatic selective pruning could provide important theoretical basis for the future pruning robot of jujube tree.This paper focused on information acquisition of jujube tree in the natural environment,and automatic recognition and length and diameter parameters extraction for pruning branches combined with deep learning method.Finally,path planning of pruning process based on random sampling algorithm was simulated,and further optimized in the virtual environment.The main research contents and conclusions of this paper were as follows:(1)A high quality 3D reconstruction method of dormant jujube tree was proposed.The quality of dormant jujube tree point cloud acquired by single Azure Kinect DK depth camera was very poor in natural environment,and so a sensing platform based on two fixed RGB-D cameras was built to obtain high quality point cloud.The calibration method of two fixed RGBD cameras was completed.The point cloud of jujube tree under three different light conditions was collected using this visual perception system.The influence of light conditions for the point cloud was analyzed after point cloud preprocessing.In order to solve the breakpoint of jujube tree branches caused by occlusion between jujube branches,a skeleton point registration algorithm was proposed to realize 3D reconstruction for jujube tree.The classical ICP algorithm,SIFT-ICP algorithm,ISS-ICP algorithm and HARRIS-ICP algorithm were compared.The experimental results showed that the proposed algorithm with higher stability had lower registration time and error.Compared with cloudy and night,the registration time of point cloud was 0.09 s,while the registration error was 0.29 mm for sunny day.The registration time of point cloud for cloudy day was between sunny day and night,and the registration error was the smallest,which was 0.11 mm.(2)The method of pruning branches recognition based on foreground image was proposed.The color point cloud of jujube tree was obtained through the visual system,and the complex background of jujube garden was removed based on the distance threshold.Meanwhile,foreground images of dormant jujube tree were constructed.Combined with the semantic segmentation models,such as Unet,PSP Net and Deep Lab V3 + based on different feature extraction networks,the semantic segmentation result of foreground images and the effects of different weather on the segmentation accuracy were analysed.The segmentation result of Deep Lab V3 + model based on Res Net-50 was the best and had good robustness.For trunk of jujube tree,the pixel accuracy and intersection over union(Io U)were 90.36% and 80.98%,while they were 80.34% and 66.69% for pruning branch,respectively.In addition,different weather conditions had no obvious influence on the segmentation accuracy for trunk and branch of jujube tree.But the identification accuracy of pruning branches needed to be further improved.(3)The method of pruning branches identification and parameter extraction based on 3D point cloud was proposed.For the reconstructed jujube trees,the deep learning method SPGNet network was used for semantic segmentation.The result of segmentation was evaluated with Class Accuracy and Io U,et al,and then the clustering algorithm was used to extract single pruning branches.The segmentation accuracy of jujube tree point cloud under different weather conditions was also analyzed.The class accuracy of the trunk was 94% and the Io U was 85%,while class accuracy of pruning branches was 83% and Io U was 75% on the test data set.Through the cluster analysis for pruning branch point clouds,45 jujube trees randomly were selected under different weather conditions.The root mean square error of pruning branch number was less than 2,which indicated that the proposed method had good robustness.Automatic extraction method for length and diameter parameters of pruning branch was proposed based on identifying single pruning branch.The relative error of length and diameter fitting for pruning branch was less than 1 cm and 2 mm,respectivly.(4)The path planning and optimization method of pruning process based on random sampling algorithm was established.Through the registration of the upper and lower position of two RGB-D cameras,the real height of jujube tree was obtained.The obtained point cloud was preprocessed to remove ground noise,and the virtual environment of jujube tree for pruning was constructed.Combined with random sampling algorithm,the collision-free path planning was carried out for pruning process of jujube tree,and the planned path was further optimized.The path planning process of four pruning branches was analyzed.The path length,calculation time and variation coefficient were used to evaluate the planning process.The proposed method could reduce some unnecessary random points and further reduce the length for the pruning process.