A detection method for synchronous recognition of string tomatoes and picking points based on keypoint detection.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Greenhouse environments present challenges for tomato picking point detection due to variable lighting, stem-background color similarity, and complex plant structures.
• Existing methods struggle with low detection accuracy, hindering automated harvesting systems.

Purpose of the Study:

• To develop an improved keypoint detection method for the synchronous recognition of string tomatoes and their picking points.
• To enhance detection accuracy and inference speed for automated agricultural applications.

Main Methods:

• A novel YOLOv8-TP model was constructed based on YOLOv8n-pose, incorporating a C2f-OREPA module for reduced computational load.
• A PSA mechanism was added after the backbone network to boost accuracy and performance.
• CGAFusion was integrated into the Neck to enhance feature extraction by adaptively emphasizing relevant features.

Main Results:

• The YOLOv8-TP model achieved 89.8% accuracy in synchronously recognizing tomatoes and picking points, with an inference speed of 154.7 FPS.
• Compared to the baseline YOLOv8n-pose, YOLOv8-TP showed improvements in precision (0.6%), mAP@.5 (1%), mAP@.5:.95 (2%), and F1-score (1%).
• The model demonstrated a Euclidean distance error of less than 25 pixels and a depth error of less than 3 millimeters, with an 8.1% reduction in model complexity.

Conclusions:

• The proposed YOLOv8-TP method offers excellent detection performance for string tomatoes and their picking points in challenging greenhouse conditions.
• This approach provides a valuable reference model for advancing automated harvesting and agricultural robotics.