Abstract
Understanding and analyzing animal behaviors is crucial for gaining profound insights into the health, needs, and overall well-being of the animal. This involves measuring and monitoring factors such as size, growth, poses, and actions. The analysis of animal behavior holds significant importance in a wide range of domains and industries, such as livestock farming, veterinary sciences, scientific research, ecological and conservation studies.
By studying the livestock animal behaviors, farmers and ranchers can enhance animal welfare, diagnose animal health issues, optimize feeding practices, and improve farm productivity. Studying the behaviors of different types of animals in the wild and various aspects of their lives, including their movements and habits, enriches our knowledge of their fascinating lives and the vast biodiversity of wildlife. This knowledge is also crucial for improving wildlife management and conservation efforts. Sometimes, such knowledge can also be important to make informed public health strategies.
Hence, given the growing interest and immense significance of the study of animals and their behaviors as well as its potential applications in a wide array of industries, researchers and practitioners have sought to develop various technologies, including computer vision and Artificial Intelligence (AI), to accurately measure animals, as well as analyze their movements and behaviors. Henceforth, this thesis focuses on the important problems in animal behavioral analysis and seeks to further advance the development of animal behavioral analysis through the use of Artificial Intelligence (AI).
First, we propose mtYOLO, a multi-task model that concurrently provides different vital characteristics (bounding box, pose information, and instance segmentation mask) of a livestock. Such an intelligent analytics system for precision livestock farming is especially useful for farms that have limited resources and computational power. Next, we present Animal Kingdom, the first in-the-wild diverse animal behavioral analysis dataset. We propose Chameleon, an animal action recognition model that capitalizes on the vast knowledge in Large Vision Language Models to enable the model to learn both generic animal actions and animal-specific actions. Lastly, while visual analysis of images and video footages can reveal a lot about animal behaviors, sometimes visual analysis is neither possible nor practical in real-life scenarios, thereby requiring us to depend on biological signals. Thus, we explore how unique bioacoustics signals can be leveraged to detect and classify dangerous species of the tiny yet lethal mosquitoes, and this shall be useful in strengthening public health strategies and efforts.
Speaker’s Profile
Ong Kian Eng is a PhD candidate at the Information Systems Technology and Design (ISTD) pillar at the Singapore University of Technology and Design (SUTD). Prior to pursuing his PhD, he obtained his Bachelor of Science (Honours) in Life Sciences (Specialization in Biomedical Sciences) and Master of Technology in Enterprise Business Analytics from the National University of Singapore.
