Abstract
Abstract
Honeybees play a vital role in the environment, particularly in pollination, yet their
populations are declining at an alarming rate. They interact not only within the colony
but also extensively with their surroundings. Understanding their behavior is crucial
for conservation efforts, and a key aspect of this is accurately estimating the pose of
individual bees. Traditional methods require attaching physical markers, such as QRcoded
tags, to the bees’ bodies, which can be intrusive. Recent advancements and
deep learning-based computer vision have enabled estimate the poses of bees which
includes location and orientation in a coordinate system.These deep learning models
use supervised learning meaning they require labeled data which includes both the
input information and corresponding annotations.They require on extensive labeled
datasets for effective learning and robustness. Moreover, these models struggle with
long-term data distribution shifts that are not represented in the training data.
This thesis presents a novel pose estimation method for honeybees that achieves comparable
accuracy to state-of-the-art methods while requiring less labeled data in the
long term. Our method requires only a small amount of labeled data for initial training
and adapts to dynamically changing scenes with different data distributions through
lifelong learning. This allows the model to systematically update its weights, ensuring
consistent accuracy over time. We demonstrate significant improvements in pose
estimation accuracy in long term, even with minimal labeled data. highlighting the
potential of our system to contribute the study and monitoring of honeybee populations.