Computer vision-based steering path visualization of headlands in soybean fields.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Autonomous steering in soybean fields faces accuracy challenges, particularly in headland areas.
• Existing autopilot systems require enhanced path planning for reliable operation during critical growth stages (V3-V8).

Purpose of the Study:

• To develop a dynamic navigation line visualization method for improving autonomous steering accuracy in soybean headland areas.
• To enhance the path planning capabilities of autopilot systems for agricultural machines.

Main Methods:

• Utilized an improved lightweight YOLO-PFL model for efficient headland detection (95.600% mAP@0.5).
• Developed a 3D positioning model using binocular stereo vision with controlled distance error (<6.000% up to 10m).
• Obtained interference-resistant crop row centerlines using HSV color space, morphological operations, and least squares fitting, addressing issues like straw and illumination changes.

Main Results:

• Achieved high precision (94.100%) and recall (92.700%) in headland detection with a lightweight model suitable for embedded systems.
• Demonstrated accurate 3D positioning and robust crop row centerline extraction with minimal orientation error (-0.473°).
• Successfully generated real-time steering paths with acceptable errors through fusion of 3D positioning and orientation data.

Conclusions:

• The proposed dynamic navigation line visualization method effectively enhances autonomous steering accuracy in soybean fields.
• This technology provides a reliable navigation reference for automated agricultural machinery, supporting intelligent precision agriculture.
• The method offers technical support for advancing smart farming equipment and practices.