Synthetic data generation for deep learning model training to understand livestock behavior

Maraghehmoghaddam, Armin
Major Professor
Joshua Peschel
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Agricultural and Biosystems Engineering

Continuous monitoring of livestock is significant in enabling the early detection of impaired and deteriorating health conditions and contributes to taking preventive measures in controlling and reducing the rate of illness or disease in livestock. Therefore, research on methods and applications for improving livestock monitoring systems in accurately and in-time detection of animal behavioral changes is of utmost importance in animal health and welfare study and practice. The emergence of new technologies provides the foundation to develop automated systems for constant livestock monitoring in farms. Among all new approaches, cameras and video recording have gained popularity due to the non-invasive platform that they offer. Increasing computational power in recent years provided a unique opportunity for applying artificial neural networks to develop models for specific tasks such as detection and classification of animals and their behaviors.

Currently, image and video analysis of livestock recordings are used as an approach for data preparation to develop detection and classification models and investigate animal behavioral changes. The process of data preparation including collection, cleaning, and labeling is prohibitively expensive, time-consuming, and laborious. An alternative to real images and videos could be using synthetically-generated visual data using which in training and developing object detectors and classifiers. However, evaluation of the feasibility of synthetically-generated visual data for training deep learning models with applications in livestock monitoring is an unexplored area of research.

Therefore, this study aims at developing a novel pipeline and platform to automate synthetic data generation and facilitate model development by eliminating the data preparation step. The objectives of the study are to: investigate the feasibility of generating and using synthetic visual data to train deep learning classifiers for object detection and classification; identify properties of synthetic data that are necessary for animal behavior characterization; and determine the best approaches for real-time analysis and detection of livestock behavioral changes using the synthetically-generated data of this study.

The study proposes approaches for generation, validation, and enhancement of synthetic data of an animal in order to address current obstacles in applying such data for object detection, which leads to developing reliable and accurate object detection models for livestock systems. Furthermore, the study provides guidelines for properly selecting deep learning object detectors, as well as methods for tuning and optimizing the performance of the models for applications in livestock monitoring.

The developed tool in this dissertation work contributes not only in reducing time, costs and labors of current data collection and analysis practices for detecting livestock behavioral changes, but also provides a solid ground for using synthetic data instead of real images for developing a reliable automated system for livestock monitoring in the field of animal science and behavior analysis.

The beneficiaries of the study include animal behavior researchers and practitioners, as well as livestock farm operators and managers. The research community can use the findings of this study to further explore the methodology of this research and develop new tools and applications based on the provided guidelines and developed framework. In addition, farm managers and operators can apply the developed tool for monitoring livestock and detect and classify animal behavioral activities to reduce or prevent livestock loss and improve animal welfare.