YOLOv5-POS: research on cabbage pose prediction method based on multi-task perception technology.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Mechanical harvesting of cabbages faces challenges due to complex plant structures, hindering accurate head and root detection.
• Existing methods struggle with precise segmentation and posture identification, leading to potential crop damage.

Purpose of the Study:

• To develop an efficient and accurate cabbage posture prediction method for automated harvesting.
• To enhance the detection and segmentation of cabbage heads and roots.
• To enable precise prediction of cabbage root growth.

Main Methods:

• Introduced YOLOv5-POS, integrating C-RepGFPN for improved feature extraction and a refined C-Seg head for segmentation.
• Developed a root growth prediction model using Bézier curves and geometric moments for key point identification.
• Utilized the anti-gravity stem-seeking principle to accurately identify root-head junctions.

Main Results:

• YOLOv5-POS achieved an F1 score of 98.8% for head/root detection and 93.5% for instance segmentation.
• Posture recognition error was low (1.38° absolute, 2.32% relative), with 98% accuracy for root growth prediction.
• Real-time processing completed within 28 milliseconds, facilitating immediate application in harvesting.

Conclusions:

• YOLOv5-POS provides a highly accurate and efficient solution for cabbage segmentation and posture prediction.
• The developed models significantly improve automated harvesting by minimizing crop damage and increasing operational efficiency.
• This approach overcomes limitations in identifying submerged root structures, offering a robust solution for agricultural robotics.