Dealing with uncertainty: Deep learning for robust animal monitoring in uncontrolled environments

Samenvatting (Engelstalig)

Monitoring animal welfare is crucial for ensuring the health and well-being of livestock. While manual monitoring is time-consuming and subjective, automated animal monitoring using cameras and AI offers the potential for continuous and objective assessment of animal welfare indicators and behavior. However, practical challenges such as high animal density, movement, poor lighting, and dust complicate reliable camera-based monitoring in barns.
Therefore, this PhD thesis focuses on developing robust AI-based methods to monitor animals under challenging conditions and integrates them into a comprehensive monitoring framework. This framework comprises a neural network for animal detection and assessment, as well as methods for estimating the uncertainty of neural network predictions, robust assessment from video sequences, and re-identification of individual chickens using deep learning.
While plumage condition in laying hens served as a representative case for the development and evaluation of the various methods, they were designed for transferability and tested on additional data.