Global path planning for navigating orchard vehicle based on fruit tree positioning and planting rows detection from UAV imagery

This paper introduces a methodology for vehicles navigation in orchard management, based on fruit trees geolocating, planting rows detection and an improved Dijkstra algorithm. A small object detection deep learning method is proposed to enhance the detection performance of fruit trees from UAV-acquired imagery map tiles, by integrating the Large Selection Kernel and Gather-Distribute feature fusion (LSK-GD) modules. A two-step plantation-row detection algorithm is established, to merge detection results from serial map tiles to global locations and cluster trees with geo-location into rows in a global view, considering the planting metrics including the spacing in-row between adjacent trees and row spacing. Based on the calculated results of planting rows, a path planning algorithm is proposed to navigate both ground and aerial orchard vehicles and perform essential management tasks in orchards. The test results show that the detection performance of LSK-GD CNN models surpasses that of other classic models. Based on the proposed methodology, the estimated tree numbers closely match the actual tree numbers (939 → 940, and 1951 → 1650). Furthermore, the estimated row numbers both remain the same as the counted ones, with a maximum angle deviation of less than 2 degrees and an average spacing deviation of less than 0.10 m. The calculated Root Mean Square Error of the automated UGV and UAV farming planning paths is less than 0.5 m. Both the calculation time and path length using the proposed method remains shorter than those using other planning methods. The overall computational times of the data mining are 27.7 and 71.02 s for two selected fields with areas of 1.34 and 2.82 acres, respectively.