ALNet: towards real-time and accurate maize row detection via anchor-line network

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Accurate crop row detection is critical for autonomous agricultural machinery navigation.
• Existing deep learning methods face challenges with computational cost, edge deployment, and balancing accuracy with speed.
• Maize row detection requires specialized approaches due to elongated geometric structures.

Purpose of the Study:

• To develop a lightweight deep learning model (ALNet) for efficient and accurate maize row detection.
• To improve the robustness of row detection under challenging field conditions.
• To enable real-time deployment of navigation systems on edge devices.

Main Methods:

• Introduced the Anchor-Line mechanism for end-to-end row detection as a regression task.
• Replaced pixel-wise convolutions with row-aligned kernel operations for reduced computation.
• Incorporated an Attention-guided ROI Align module with a Dual-Axis Extrusion Transformer (DAE-Former) for enhanced feature interaction.
• Developed a Row IoU (RIoU) loss function to improve localization accuracy.

Main Results:

• ALNet achieved an mF1 score of 59.60, outperforming competing methods by over 9.24 points.
• Demonstrated a high inference speed of 161.26 FPS with a low computational cost of 11.9 GFlops.
• Showcased robustness in challenging conditions like weed infestation, low light, and wind distortion.

Conclusions:

• ALNet offers a practical and scalable solution for intelligent visual navigation in precision agriculture.
• The lightweight design and high efficiency make ALNet suitable for real-time edge deployment.
• The proposed Anchor-Line mechanism and attention modules significantly advance crop row detection capabilities.