Instance Segmentation with Mask R-CNN Applied to Loose-Housed Dairy Cows in a Multi-Camera Setting

Area of Science:

• Agricultural Engineering
• Animal Science
• Computer Vision

Background:

• Increasing dairy herd sizes necessitate automated health and welfare monitoring systems.
• Dairy cow social structure significantly impacts animal welfare.
• Human observation can alter animal behavior, making automated detection systems beneficial.

Purpose of the Study:

• To develop and evaluate an automated camera-based system for detecting dairy cows.
• To analyze dairy cattle herd activity and resource utilization in loose-housing environments.
• To assess the impact of training data set size on model performance.

Main Methods:

• Utilized eight surveillance cameras in a loose-housing barn with 36 lactating Holstein Friesian cows.
• Trained a Mask R-CNN convolutional neural network model for pixel-level cow segmentation.
• Employed transfer learning with a pre-trained model on Microsoft COCO dataset and annotated recordings.

Main Results:

• Achieved 91% average precision for bounding box detection (IOU = 0.5).
• Achieved 85% average precision for segmentation mask detection (IOU = 0.5).
• Demonstrated a relationship between training data set size and model performance.

Conclusions:

• The Mask R-CNN model provides a robust technical foundation for automated dairy cattle monitoring.
• Automated detection facilitates objective analysis of herd activity and resource management.
• This technology supports improved animal welfare assessment in large herds.