Determining the posture and location of pigs using an object detection model under different lighting conditions

Area of Science:

• Animal Science
• Computer Vision
• Agricultural Technology

Background:

• Computer vision, particularly object detection, is increasingly used for monitoring pig behavior, including location and posture.
• Lighting conditions significantly impact the performance of detection models and can influence pig activity and resting patterns.
• Accurate pig behavior monitoring is crucial for welfare assessment and optimizing farming practices.

Purpose of the Study:

• To validate a YOLOv8 object detection model for identifying pig postures (standing, sitting, sternal lying, lateral lying) under diverse lighting conditions.
• To assess the model's performance across varying light intensity, spectrum, and uniformity.
• To evaluate the model's accuracy in localizing and classifying pig behaviors in group-housed settings.

Main Methods:

• Development and training of a YOLOv8 object detection model using annotated data from pigs aged 10-24 weeks.
• Inclusion of 10 different lighting settings (varying intensity, spectrum, uniformity) in the training, validation, and test datasets.
• Evaluation of detection accuracy using mean average precision (mAP), precision, sensitivity, and F1 scores, with localization accuracy measured by intersection over union (IoU).

Main Results:

• The model achieved a mean average precision (mAP) of 89.4% for pig detection across different lighting conditions.
• Standing posture detection was most accurate; sitting posture detection had the lowest sensitivity and F1 score due to visual confusion and dataset limitations.
• Grouping sternal and lateral lying postures improved detection (mAP = 97.0%), and localization accuracy (IoU) exceeded 95.5% for 75% of the dataset.

Conclusions:

• The YOLOv8 model demonstrates robust performance in detecting pig location and postures under varied lighting, with minor challenges in specific conditions like warm or uneven light.
• Distinguishing between specific lying postures and sitting can be improved by data augmentation and by analyzing active versus resting behaviors collectively.
• The study confirms the potential of computer vision for reliable pig behavior monitoring, though challenges in individual pig tracking and occlusion require further research.