Research On Key Technologies Of Autonomous Navigation Between Rows In Orchard

From 1990 to 2019,Chinese fruit production ranked first in the world,and fruit industry has gradually become an important pillar of the rural economy.In order to further improve production efficiency and fruit quality,intelligent agricultural equipment is an inevitable trend of future development.The equipment can use sensors to perceive environment and its own state,and complete predetermined tasks in the orchard environment.As the key technology of agricultural intelligence,autonomous navigation between rows of orchards can realize autonomous operation between rows,reducing labor intensity and improving work efficiency.Orchard is a complex non-structural unknown environment.The key to achieving autonomous operation is environment perception,pose estimation and autonomous positioning.This paper focuses on monocular vision navigation and lidar navigation,(1)and(2)belong to lidar navigation,and(3)-(5)belong to monocular vision navigation.The main work and innovations are as follows:(1)For the problem of filter sensitivity to Li DAR point cloud density,an adaptive radius filtering method is proposed.This method dynamically calculates the filter radius according to the target distance to realize multi-scale noise removal.The experimental results show that compared with the raw data,the DBSCAN precision rate is increased by 0.40,the recall rate is increased by 0.34,and the average calculation time is 43 ms.(2)For the problem of road extraction and pose estimation based on Li DAR,a passable area extraction method based on intersecting pixels is proposed.The experimental results show that the method can accurately extract the passable area,yaw angle and lateral displacement in both dense planting mode and non-dense planting mode,and has high adaptability and stability.(3)Scene semantic information extraction technique is studied,based on the improved Mask R-CNN network,realize the segmentation of roads and tree trunks,and calculate the boundary equation and vanishing point coordinates based on the Hough transform.Experimental results show that this method extracts multi-scale target information in complex lighting conditions,and provides an effective reference for monocular pose estimation and visual navigation.(4)For the problem of high-precision pose estimation,a geometric imaging model of roads between rows of orchards is established.Based on this model,the calculation methods of yaw angle,lateral offset and road width are realized.The experimental results show that the measurement accuracy of road width is 0.989,the average error of yaw angle is 0.042,the average error of lateral displacement is 0.048,and the average calculation time is 56 ms.(5)For the problem of fruit tree positioning,the relative position of fruit trees is calculated based on the monocular camera by using the collinearity of the tree rows and the parallelism of the boundaries.The experimental results show that the average error of horizontal positioning is 0.038,the average error of longitudinal positioning is0.027,the average error of plant spacing estimation is 0.081,and the average calculation time is 1.1ms.