Cattle behavior recognition from accelerometer data: Leveraging in-situ cross-device model learning

Abstract

Automating livestock behavior recognition using wearable sensors offers significant benefits for monitoring animal health, ensuring welfare, and enhancing farm productivity. While collar-mounted accelerometers provide useful data leading to accurate behavior recognition models, ear-tags offer greater practicality and scalability. However, ear-tag data is affected by independent ear movements (e.g., for flicking flies), necessitating extensive labeled data for accurate recognition, which is time-consuming and costly to obtain. To address this challenge, we propose a pioneering cross-device learning approach. By leveraging a pre-trained behavior recognition model from collar data to guide ear-tag model training, we significantly reduce the need for manual labeling of ear-tag data. This facilitates the development of efficient and scalable behavior recognition models suitable for wider deployment. Additionally, we introduce a novel deep neural network architecture that integrates linearly-constrained convolutional layers specifically designed to counteract gravity’s impact on accelerometer data, along with a confidence penalty term to combat overfitting when training on limited labeled data. Evaluation on real-world cattle data demonstrates that our approach yields ear-tag model performance nearly on par with collar models. This breakthrough paves the way for personalized behavior recognition models using ear-tags, requiring only brief periods of collar-based labeling per animal.