Improved Multi-Size, Multi-Target and 3D Position Detection Network for Flowering Chinese Cabbage Based on YOLOv8

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Autonomous harvesting in agriculture requires precise plant detection and localization.
• Challenges include dense planting, small targets, and occlusions in natural field conditions.
• Accurate maturity and 3D positioning of flowering Chinese cabbage (Brassica rapa var. chinensis) are crucial for robotic harvesting.

Purpose of the Study:

• To develop an improved object detection and 3D positioning system for flowering Chinese cabbage.
• To enhance the YOLOv8 network for better feature extraction, multi-scale fusion, and small object detection.
• To integrate real-time tracking and depth data for accurate field localization.

Main Methods:

• Proposed YOLOv8-Improved network with C2F-MLCA module for spatial feature extraction.
• Incorporated P2 detection layer and BiFPN for enhanced multi-scale feature fusion.
• Utilized Wise-IoU and Inner-IoU as a novel loss function for optimized detection.
• Integrated ByteTrack for video tracking and RGB-D camera for 3D positioning.

Main Results:

• YOLOv8-Improved achieved 86.5% precision and 86.0% recall for maturity detection.
• mAP50 and mAP75 improved by 2.9% and 4.7% over the original YOLOv8 network.
• The improved network reduced parameters by 25.43% while maintaining high performance.

Conclusions:

• The enhanced YOLOv8 algorithm demonstrates robust and real-time detection capabilities for flowering Chinese cabbage.
• This system provides significant technical support for automated harvesting management in unstructured farm environments.
• The improvements in feature extraction, fusion, and loss function contribute to superior detection accuracy.