Automatic Apple Detection and Counting with AD-YOLO and MR-SORT.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate fruit counting is crucial for agricultural production management, impacting yield estimation and decision-making.
• Existing tracking-by-detection algorithms struggle with occlusion and lighting variations in orchards, hindering automatic and precise apple counting.

Purpose of the Study:

• To develop an advanced video-based multiple-object tracking method for accurate fruit counting in complex agricultural environments.
• To enhance the performance of object detection and tracking algorithms for improved orchard yield estimation.

Main Methods:

• Proposed AD-YOLO model integrating Omni-dimensional Dynamic Convolution (ODConv), Global Attention Mechanism (GAM), and Soft Spatial Pyramid Pooling Layer (SSPPL) for improved detection.
• Developed an enhanced BoT-SORT algorithm by incorporating a verification mechanism, SURF feature descriptors, and Vector of Local Aggregate Descriptors (VLAD) for robust tracking.
• Utilized video-based multiple-object tracking for real-time fruit counting in orchard settings.

Main Results:

• The AD-YOLO model achieved a 3.1% higher mAP (96.4%) compared to the standard YOLOv8.
• The improved tracking algorithm reduced ID switches by 35.6% (297 fewer).
• Achieved 85.6% multiple-object tracking accuracy, an average counting error of 0.07, and an R² of 0.98.

Conclusions:

• The MR-SORT method significantly improves the accuracy and reliability of automatic fruit counting in orchards.
• The enhanced detection and tracking capabilities offer a robust solution for agricultural yield estimation and management.
• The proposed approach demonstrates potential for accurate counting of various fruit types beyond apples.