GPS-free Autonomous Navigation in Cluttered Tree Rows with Deep Semantic Segmentation

TL;DR

The paper addresses GPS-denied autonomous navigation in dense crop rows by introducing SegMin and SegMinD, two segmentation-based controllers that rely on RGB-D input and a deep segmentation model trained exclusively on synthetic data. SegMin uses a column-wise histogram of depth-filtered segmentation to locate the row center, while SegMinD adds depth-based weighting to emphasize near vegetation, improving performance in wide or curved rows. A MobileNetV3-based segmentation network with LR-ASPP head is trained on the AgriSeg synthetic dataset and validated on real crops, achieving robust IoU across vineyards, pears, and apples. Extensive simulations in Gazebo and real-field tests demonstrate that SegMin/SegMinD provide GPS-free, robust navigation through tall canopies, outperforming a SegZeros baseline and enabling practical deployment in diverse agricultural settings.

Abstract

Segmentation-based autonomous navigation has recently been presented as an appealing approach to guiding robotic platforms through crop rows without requiring perfect GPS localization. Nevertheless, current techniques are restricted to situations where the distinct separation between the plants and the sky allows for the identification of the row's center. However, tall, dense vegetation, such as high tree rows and orchards, is the primary cause of GPS signal blockage. In this study, we increase the overall robustness and adaptability of the control algorithm by extending the segmentation-based robotic guiding to those cases where canopies and branches occlude the sky and prevent the utilization of GPS and earlier approaches. An efficient Deep Neural Network architecture has been used to address semantic segmentation, performing the training with synthetic data only. Numerous vineyards and tree fields have undergone extensive testing in both simulation and real-world to show the solution's competitive benefits.

Figures (9)

• 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 left picture, a real vineyard row on the right.
• 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: Contrasting the histograms produced by the two distinct algorithms, considering the RGB frame on the right, reveals that SegMinD provides a more defined and less ambiguous global minimum point.
• Figure 4: The Deep Neural Network utilized in this study features a backbone of MobileNetV3 and an LR-ASPP head, as detailed in howard2019searching. The spatial scaling factor of the features in comparison to the input size is provided beneath each block.
• Figure 5: Test of semantic segmentation DNN on real-world test samples from vineyard (top), pear trees (middle) and apple trees (bottom) fields. For each crop, RGB input image (left), ground truth mask (center) and the predicted mask (right) are reported.