Continuous Picking Path Planning Based on Lightweight Marigold Corollas Recognition in the Field.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Automated harvesting requires precise object recognition and efficient path planning.
• Complex natural conditions like illumination and occlusion pose significant challenges.

Purpose of the Study:

• To develop an integrated lightweight framework for marigold corolla recognition and path planning.
• To enhance the accuracy and efficiency of automated marigold harvesting in unstructured environments.

Main Methods:

• Developed MPD-YOLO, an optimized YOLOv11n model with a Multi-scale Information Enhancement Module (MSEE).
• Implemented structured pruning and knowledge distillation for model compression and accuracy recovery.
• Utilized an adaptive ant colony algorithm with dynamic parameter adjustment for path planning.

Main Results:

• MPD-YOLO achieved high precision (P: 89.8%, mAP@0.5: 95.1%) and robustness, with a 6.8% recall increase.
• Model size reduced to 2.1 MB (39.6% of original) with a computational load of 3.2 GFLOPs.
• Path planning time reduced to 2.2 s (68.6% speedup) using the adaptive ant colony algorithm.

Conclusions:

• The integrated framework significantly improves perception accuracy and operational efficiency for automated marigold harvesting.
• The lightweight and robust MPD-YOLO model provides a viable solution for real-world agricultural applications.
• This research offers robust technical support for continuous automated operations in challenging agricultural settings.