Lightweight Detection and Adaptive Path Planning for Selective Hotan Rose Harvesting.

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Selective harvesting of Hotan roses necessitates distinguishing buds from blooms for industrial applications.
• Balancing high detection accuracy with computational efficiency for edge deployment is a significant challenge in automated harvesting systems.

Purpose of the Study:

• To develop an integrated framework for automated selective rose harvesting.
• To enhance detection accuracy and computational efficiency for edge devices.
• To optimize path planning for efficient robotic harvesting operations.

Main Methods:

• Proposed a lightweight detection model, Rose_YOLO, optimizing YOLOv8n with C2f-Faster-CGLU and a Rose_Head.
• Developed the ROSE algorithm, integrating a genetic algorithm (GA) with reciprocating search for adaptive path planning.
• Evaluated model performance on detection accuracy, parameter count, model size, computational complexity, path length, and runtime.

Main Results:

• Rose_YOLO achieved high precision (90.4% blooms, 88.4% buds) and mAP@0.5 (96.6% blooms, 91.7% buds).
• The model significantly reduced parameters (47.46%), size (3.19 MB), and computational complexity (4.6 GFLOPs) compared to YOLOv8n.
• The ROSE algorithm generated significantly shorter paths (73.1% shorter than reciprocating, 51.6% shorter than GA) with rapid execution (7.33 ms runtime).

Conclusions:

• The integrated framework offers a superior balance between lightweight design and detection performance for automated rose harvesting.
• Successful edge deployment demonstrates effectiveness in real-time visual guidance and efficient path planning.
• Presents a robust technical solution for automated selective rose harvesting in complex field environments.