Doctoral Dissertations
Date of Award
12-2025
Degree Type
Dissertation
Degree Name
Doctor of Philosophy
Major
Biosystems Engineering
Major Professor
Hao Gan
Committee Members
Robert Burns, Lannett Edwards, Hairong Qi
Abstract
This study investigates the use of computer vision as a non-invasive, scalable approach for monitoring key physiological, behavioral, and physical indicators of beef cattle health and welfare. The main goal was to develop automated, camera-based frameworks capable of replacing or augmenting manual or contact-based monitoring methods commonly used in livestock management.
Three integrated systems were developed and validated under real farm conditions: (1) breathing-rate detection, (2) drinking-behavior estimation, and (3) body condition scoring (BCS). For respiration monitoring, a learning-based motion magnification and pose-estimation pipeline was implemented to amplify low-flank movements associated with breathing. Adaptive signal analysis and a hybrid winner strategy were applied to extract breathing frequencies from motion signals. The model achieved a mean bias of 1.84% and limits of agreement within ±20% when compared with manually counted respiration, with accuracy remaining stable across lighting conditions and animal variations.
The drinking-behavior framework utilized pose estimation to track head and neck keypoints and employed Long Short-Term Memory (LSTM) and Transformer networks to classify drinking events from video sequences. The LSTM achieved 95.5% accuracy and an F1-score of 0.97. Behavioral features derived from the classified events, including duration and frequency, were integrated with water flow–sensor data to predict daily water intake (R² = 0.81). Results revealed increased water consumption during estrus periods, highlighting the biological relevance of behavioral metrics derived from vision data.
The automated BCS framework combined multi-view image analysis and feature extraction from anatomically significant regions such as the spine, tail head, and whole body, identified by a YOLO detector. Features such as tail-head angle, spine ridge visibility, fat-knob counts, and contour roundness were used in classification models, achieving up to 0.93 accuracy with Random Forests.
Collectively, these three studies demonstrate that camera-based monitoring can provide continuous, objective, and interpretable assessment of cattle physiology, behavior, and condition. The findings establish a foundation for integrated, AI-driven livestock management systems capable of supporting early disease detection, welfare assessment, and data-driven decision-making. This research advances precision livestock farming toward autonomous, welfare-focused, and sustainable beef production.
Recommended Citation
Islam, Md Nafiul, "Smart Cattle Monitoring Using Computer Vision to Detect Physiological, Behavioral, and Physical Characteristics. " PhD diss., University of Tennessee, 2025.
https://trace.tennessee.edu/utk_graddiss/13605