Autonomous Navigation in Rows of Trees and High Crops with Deep Semantic Segmentation

TL;DR

This paper tackles GPS-denied autonomous navigation through dense crop rows by leveraging RGB-D semantic segmentation to identify a row center without relying on sky visibility. It introduces two variants, SegMin and SegMinD, which convert the segmentation mask into a column-wise histogram and select a central minimum to steer the robot, with SegMinD incorporating depth weighting to handle wide or tall canopies. The segmentation backbone is MobilenetV3 with an LR-ASPP head, trained on AgriSeg for real-time performance, and the control law translates the histogram minimum into linear and angular velocities with EMA smoothing. Extensive Gazebo simulations across vineyards, pergola layouts, pear fields, and high-tree rows demonstrate that SegMin and SegMinD outperform a prior SegZeros baseline, offering improved robustness and navigation efficiency in challenging canopy conditions. The approach shows potential for practical deployment in GPS-denied agricultural environments and sets the stage for real-world validation and hardware-level optimizations.

Abstract

Segmentation-based autonomous navigation has recently been proposed as a promising methodology to guide robotic platforms through crop rows without requiring precise GPS localization. However, existing methods are limited to scenarios where the centre of the row can be identified thanks to the sharp distinction between the plants and the sky. However, GPS signal obstruction mainly occurs in the case of tall, dense vegetation, such as high tree rows and orchards. In this work, we extend the segmentation-based robotic guidance to those scenarios where canopies and branches occlude the sky and hinder the usage of GPS and previous methods, increasing the overall robustness and adaptability of the control algorithm. Extensive experimentation on several realistic simulated tree fields and vineyards demonstrates the competitive advantages of the proposed solution.

Figures (6)

• Figure 1: The proposed SegMin and SegMinD algorithms allow to precisely guide an autonomous mobile robot through a dense tree row solely using an RGB-D camera. A pear crop row in Gazebo is shown in the picture.
• Figure 2: Scheme of the overall proposed navigation pipeline. The RGB image is fed into the segmentation network, thus the predicted segmentation mask $\hat{\textbf{X}}_{seg}^{t}$ is refined using the depth frame to obtain $\hat{\textbf{X}}_{segDepth}^{t}$. The blue arrow refers to the SegMin variant, and red arrows refer to the SegMinD variant to compute the sum histogram over the mask columns. Images are taken from navigation in the tall trees simulation world.
• Figure 3: Sample outputs of the proposed SegMinD algorithm for High Trees (a), Pear Trees (b), Pergola Vineyard (c), and Vineyard (d). Predicted segmentation masks are refined cutting values exceeding a depth threshold. The sum over mask columns provide the histograms used to identify the centre of the row as its global minimum.
• Figure 4: Gazebo simulated environments used to test the SegMin approach in relevant different crops rows: wide rows composed of high trees (a), a narrow pear tree row (b), a pergola vineyard with asymmetric rows (c), straight and curved vineyard rows (d). In the last case, the tests were carried out in the second row from above and the second row from below.
• Figure 5: Comparison of the two histograms obtained using the two different algorithms, given the RGB frame on the right. It can be noticed how SegMinD offers a narrower and less ambiguous global minimum point.