CTDA: an accurate and efficient cherry tomato detection algorithm in complex environments.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic harvesting of cherry tomatoes faces challenges from lighting, occlusion, and overlapping fruit.
• Accurate and efficient detection is crucial for successful automated harvesting in unstructured environments.

Purpose of the Study:

• To propose a precise, real-time, and robust target detection algorithm (CTDA) for cherry tomato harvesting.
• To enhance the accuracy and efficiency of robotic vision systems in complex natural harvesting conditions.

Main Methods:

• The CTDA model is based on YOLOv8, featuring a restructured backbone with lightweight downsampling and adaptive weights.
• It incorporates SoftPool in SPPF (SPPFS) for efficient feature utilization and multi-scale fusion.
• An attention-driven dynamic head enhances feature capture across scales for improved recognition.

Main Results:

• CTDA achieved 94.3% detection accuracy, 91.5% recall, and 95.3% average precision.
• The model demonstrated a mAP@0.5:0.95 of 76.5% and a speed of 154.1 FPS.
• Compared to YOLOv8, CTDA improved mAP by 2.9% with a smaller model size (6.7M).

Conclusions:

• The CTDA model is effective for cherry tomato detection in complex environments, showing robustness to lighting variations and occlusion.
• It supports rapid detection on edge devices, providing a strong foundation for automated cherry tomato picking.
• The algorithm enhances adaptability for dense, small target scenarios, crucial for agricultural applications.