IYOLO-FAM: Improved YOLOv8 with feature attention mechanism for cow behaviour detection

Ahmad, Misbah; Zhang, Wenhao; Smith, Melvyn; Brilot, Ben; Bell, Matt

Abstract

We introduced IYOLO-FAM (Improved YOLOv8 with Feature Attention Mechanism) for detecting cow behaviours. By leveraging the robust YOLOv8 architecture improved with Feature Attention Mechanisms (FAM), Squeeze-and-Excitation (SE) blocks and data augmentation techniques, we enhanced the ability of the model to focus on salient features and generalize across a diverse farm environment. The experimental results demonstrated that IYOLO-FAM outperforms baseline YOLO models, achieving a mean Average Precision (mAP) of 88% at an IoU threshold of 0.5 and 70% across IoU thresholds from 0.5 to 0.95. These results highlighted substantial improvements over previous versions, particularly in detecting specific cow behaviours such as eating, lying, standing, and walking. The integration of SE blocks and FAM within the YOLOv8 framework proved effective in highlighting relevant features and enhancing detection accuracy, underscoring the significance of integrating advanced deep learning techniques with robust data augmentation techniques to tackle the challenges posed by a real-world farm environment. The proposed approach has the potential to benefit animal welfare in real-world applications, with future research focusing on integrating multimodal data. Additionally, real-world trials will validate the model’s robustness and effectiveness in a practical farm environment.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON)
• Start Date: Oct 17, 2024
• End Date: Oct 19, 2024
• Acceptance Date: Sep 19, 2024
• Online Publication Date: Nov 20, 2024
• Publication Date: Nov 20, 2024
• Deposit Date: Oct 22, 2024
• Publicly Available Date: Dec 21, 2024
• Peer Reviewed: Peer Reviewed
• ISBN: 9798331540913
• DOI: https://doi.org/10.1109/UEMCON62879.2024.10754666
• Public URL: https://uwe-repository.worktribe.com/output/13310924