Core Concepts
A novel approach for GPS-free autonomous navigation in tall and dense crop rows using deep semantic segmentation and depth information to estimate the center of the row.
Abstract
This work presents two variants of a novel approach for segmentation-based autonomous navigation in tall and dense crop rows, designed to tackle challenging scenarios where GPS signals are blocked by thick vegetation canopies.
The key highlights and insights are:
The proposed SegMin and SegMinD algorithms use deep semantic segmentation and depth information to estimate the center of the crop row, without relying on GPS localization.
SegMin refines the previous histogram-based approach by searching for the global minimum in the column-wise sum of the segmentation mask. SegMinD further incorporates depth information to enhance the algorithm's ability to discern the row direction.
The segmentation model is trained using only synthetic data from the AgriSeg dataset, demonstrating its generalization capabilities to real-world scenarios.
Extensive testing in simulation across diverse crop fields, including tall trees, pear orchards, pergola vineyards, and straight/curved vineyards, shows the proposed methods outperform the previous state-of-the-art.
Real-world experiments on apple, pear, and vineyard fields validate the effectiveness and robustness of the navigation algorithms in handling challenging terrain and vegetation conditions.
Stats
The simulation environments had the following dimensions:
Rows distance: 1.8-7.0 m
Plant distance: 1.0-5.0 m
Plant height: 2.0-12.5 m