Efficient Object Tracking in Unlabeled Videos for Animal Behavior Studies Using Machine Learning

Abstract

This work seeks to explore the development of a semi-supervised framework for animal tracking in controlled environments, leveraging principles from computer vision, machine learning, and deep learning. Initially conceptualized for single-object tracking, the approach will be generalized to accommodate multiple entities. Addressing the limitations posed by sparse labeled data, the method will involve user input to initialize object selection, enabling automated tracking across video sequences. Various techniques, including template matching, optical flow, and data augmentation, will be incorporated. The model's potential will be evaluated in scenarios involving complex movement patterns, with the aim of extracting relevant behavioral insights.