Adaptive Multi-ROI Agricultural Robot Navigation Line Extraction Based on Image Semantic Segmentation.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning for Agriculture

Background:

• Sustainable food production in smart agriculture relies on automated robots with precise navigation systems.
• Existing visual navigation systems face challenges from complex backgrounds, weeds, and lighting variations in field environments.
• Accurate crop row detection is crucial for agricultural robot path planning and task execution.

Purpose of the Study:

• To address limitations in visual navigation for agricultural robots by proposing an optimized deep learning model.
• To enhance the accuracy and efficiency of crop row prediction in diverse field conditions.
• To provide a robust navigation solution for intelligent agricultural robots.

Main Methods:

• Developed a Faster-U-net model, an optimized version of U-net with retained feature jump connections.
• Trained the model on a corn dataset, then used its weights for pre-training on cucumber, wheat, and tomato datasets.
• Employed B-spline curve fitting for generating navigation lines and robot yaw angles based on predicted crop ridges.

Main Results:

• The Faster-U-net model achieved high recognition accuracy (MIoU) for maize (93.86%), tomatoes (94.01%), cucumbers (93.14%), and wheat (89.10%).
• Model parameters were reduced by 65.86% with an mPA of 97.39%, demonstrating significant optimization.
• The system exhibited strong robustness across different crops with low average angle differences for navigation lines (e.g., 0.526° for cucumbers).

Conclusions:

• The proposed Faster-U-net model offers a robust and efficient solution for agricultural robot visual navigation.
• The method effectively handles complex environmental factors, improving navigation line and yaw angle prediction accuracy.
• This research provides valuable technical support for developing advanced intelligent agricultural robot navigation equipment.