Generalized perception of tree-row with distribution-peak feature in 3D point cloud for various orchards navigation

Abstract

Universal navigation is crucial for enhancing the environmental adaptability of agricultural robots, promoting large-scale manufacturing and widespread adoption of hardware, and increasing the utilization rate of agricultural robots. Autonomous navigation perception in orchards faces challenges such as the dense growth of branches and leaves obstructing key features, dynamic environmental changes, and significant structural differences across various orchard types. To achieve the goal of autonomous navigation and perception for agricultural robots across various types of orchards. In this study, we analyzed the relationship between the distribution of tree-row point cloud in LiDAR coordinate space and heading, and extracted a commonality distribution-peak feature across various orchards to broaden the generalization of the tree-row perception. Then, we addressed the impact of interference point clouds, local ground unevenness, and large heading offset on perception, and developed a generalization tree-row perception method based on distribution-peak to achieve inter-row localization task in various orchards. Experiments were conducted to validate the algorithm in several orchards of different types, sizes and seasons. Experiments were performed in multiple orchards of different types and specifications, and the results indicated that the heading mean absolute error (MAE) was from 0.88° to 1.25° and the lateral MAE was from 3.57 cm to 7.99 cm of the generalization tree-row perception method in different orchards, which meet the localization requirements for orchard navigation. This study can offer insights into the generalization of environmental perception for orchard navigation.