Development of a Machine stereo vision-based autonomous navigation system for orchard speed sprayers

Abstract

In orchards, radio frequency scintillation caused by high vegetation density can hinder the effectiveness of machinery auto guidance based on the Global Navigation Satellite System (GNSS). Signal interference not only leads to poor quality of machine operation but can also pose operational risks. In this work, we propose an alternative trajectory positioning system to supplement traditional GNSS-based navigation for machinery used in orchards. The aim of this research was to develop a deep learning machine stereo vision guidance system onboard an orchard speed sprayer. The developed system combines a collision avoidance methodology along with deep learning-driven machine vision for interrow positioning and a dead reckoning set of rules for alternating U-turns. The developed methodology was tested in 4 rows of an artificial orchard. The results show that it is possible for the embedded EfficientDet target detection algorithm to guide the equipment at 1, 1.5 and 2 km × h⁻¹ with minimum average root-mean-square errors (RMSEs) of 0.24 m, 0.20 m, and 0.31 m, respectively. When the system navigation was performed using YOLOv7, the minimum average RMSEs for each row were 0.40 m, 0.48 m, and 0.43 m, respectively, for the abovementioned speeds. The U-turn by dead reckoning showed minimum average RMSE values of 0.56 m, 0.22 m, and 0.35 m for row navigation based on EfficientDet at 1, 1.5 and 2 km × h^−1, respectively. For YOLOv7-based navigation in rows, the minimum average RMSE values at these speeds were 0.35 m, 0.81 m, and 0.44 m, respectively. This study contributes to the field by proposing an alternative navigation system for orchard machines that operates without the limitations associated with GNSS. In addition, our proposed guidance methodology introduces an RGB-D collision avoidance system with a demonstrated safety capacity for navigation under real scenario conditions.