Objective The objective of this study is to solve the problem of large longitudinal spacing of fruit trees and the inavailability of global navigation satellite system (GNSS) signals in closed-canopy orchard environment.

Method A navigation method based on three-dimensional light detection and ranging (3D LiDAR) was proposed, taking the wheeled spray robot as the research platform and the canopy mango orchard as the experimental environment. For laser point cloud preprocessing, mounting error calibration of the liDAR was initially conducted. Terrain compensation for 3D LiDAR point cloud positions was implemented via an attitude and heading reference system (AHRS). The cloth simulation filter (CSF) was employed to extract ground points. An improved statistical filtering method based on the Euclidean distance of point clouds was used to both remove noise point clouds and retain distant fruit tree point clouds. Based on the scanning characteristics of 3D LiDAR point cloud and the triangular inequality condition, an adaptive distance threshold calculation method with clustering body center constraint was designed, and the obtained body center position was projected to the X-Y plane of the navigation coordinate system to obtain the clustered body center position of the trunk point cloud. Newton’s interpolation method was used to interpolate the body-centered position data, and the random sample consensus (RANSAC) algorithm was used to fit the navigation path, i.e., NIL-RANSAC, after the interpolation was completed. In order to verify the accuracy and reliability of navigation path extraction, two methods, least squares method (LSM) and RANSAC, were used to obtain the navigation path directly and conduct comparative experiments. A linear quadratic regulator (LQR) was used for path following control.

Result Using CSF in closed-canopy orchard effectively removed weeds and uneven ground point clouds and the treatment time was only 0.03 s. The success rate of Euclidean clustering with the adaptive distance threshold within 15 m was more than 95%. LQR realized path following control, and the maximum lateral deviations of NIL-RANSAC, RANSAC and LSM were 0.26, 0.32 and 0.42 m, respectively, and the standard deviation of NIL-RANSAC was the minimum, being only 0.09 m. The navigation accuracy of the NIL-RANSAC path fitting method was better than those of RANSAC and LSM, and the average time of the complete navigation algorithm was less than 100 ms.

Conclusion The NIL-RANSAC method can meet the requirements of accurate and real-time navigation of closed-canopy orchard environment, and provide a reference for autonomous navigation of orchard ground equipment.