Real Time Pear Fruit Detection and Counting Using YOLOv4 Models and Deep SORT

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate fruit counting is crucial for yield estimation and orchard management.
• Existing methods often lack real-time capabilities or require complex hardware.

Purpose of the Study:

• To develop a robust, real-time pear fruit counter for mobile applications.
• To evaluate different variants of the YOLOv4 object detection model and the Deep SORT tracking algorithm.
• To provide a methodology for selecting optimal models in agricultural science applications.

Main Methods:

• Utilized RGB data with YOLOv4 object detection variants (YOLOv4-CSP, YOLOv4-tiny, YOLOv4) and Deep SORT multiple object tracking.
• Systematically assessed models based on accuracy (AP@0.50), speed (FPS), and computational cost (FLOPS).
• Compared two counting methods within Deep SORT: unique ID and Region of Interest (ROI) line.

Main Results:

• YOLOv4-CSP achieved the highest accuracy (AP@0.50 of 98%).
• YOLOv4-tiny demonstrated superior speed (>50 FPS) and lower computational cost (6.8-14.5 FLOPS).
• The standard YOLOv4 model balanced accuracy and real-time performance (≥24 FPS).
• The unique ID counting method with Deep SORT proved more reliable (F1count of 87.85%) due to YOLOv4's low false negative rate.

Conclusions:

• A real-time pear fruit counter using YOLOv4 and Deep SORT is feasible for mobile applications.
• The choice of YOLOv4 variant depends on specific application needs (accuracy vs. speed/cost).
• The unique ID method in Deep SORT is recommended for reliable counting in this context.