Deep Learning Models to Predict Finishing Pig Weight Using Point Clouds.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Visual assessment of farm animals for marketing is subjective and relies on caretaker skill.
• Accurate, real-time weight monitoring is crucial for animal marketing, health, and well-being assessment.

Purpose of the Study:

• To develop and evaluate a 3D Convolutional Neural Network (CNN) method for predicting farm animal weight from 3D point clouds.
• To compare the performance of the 3D CNN model with a traditional volume-based weight estimation method.

Main Methods:

• Captured 3D videos of 249 pigs (20-120 kg) using an Intel Real Sense D435 camera.
• Extracted point clouds from videos and applied the PointNet framework for weight prediction modeling.
• Compared 3D CNN model performance against a volume calculation method.

Main Results:

• The 3D CNN model achieved a high coefficient of determination (R² = 0.94) for weight prediction.
• The model demonstrated a test Root Mean Square Error (RMSE) of 6.88 kg.
• The model performed best for predicting the weight of pigs below 55 kg.

Conclusions:

• 3D deep learning on point sets shows significant potential for accurate farm animal weight prediction.
• The developed 3D CNN method is a viable tool for real-time weight estimation in livestock.
• Further research with larger datasets is recommended for optimizing prediction accuracy.