Multiple behaviour recognition of free-range broilers in cross-domain scenarios using MCA-YOLOv5

Abstract

Precision livestock farming is the trend of the future, and the use of computer vision to replace traditional manual monitoring is the key to real-time poultry disease monitoring. Behaviour changes in broilers serve as important indicators of their health status. In previous studies, most broiler research has focused on basic behaviours such as drinking, feeding and walking. However, with the transition to cage-free housing, more natural behaviours need to be monitored for welfare assessment. To address this issue, the article proposes a multi-behaviour monitoring model for broilers that combines computer vision. To improve the model accuracy without increasing the overall parameter amount. First, the Transformer-based MobileVitV3 lightweight network structure had been introduced into the YOLOv5. Then, the Coordinate Attention (CA) attention mechanism had been incorporated into the Backbone to focus on important features. Finally, a new loss function called Focal-DIOU had been innovatively proposed. The experimental results demonstrate that our proposed MCA-YOLOv5 model achieves the highest recognition accuracy, with precision (P) and mean Average Precision (mAP) reaching 90 % and 91 % respectively. The method further improved algorithm performance and reduced computational cost compared with the state-of-the-art, such as YOLOv5s, SSD, Faster-RCNN, YOLOv3, YOLOv4, and PPYOLO. Compared with these algorithms, the P of the proposed improved model increased by 2.0 %, 15.5 %, 9.0 %, 32.5 %, 2.7 %, 2.8 %, and 7.5 %, respectively. By utilising the MCA-YOLOv5 model in cross-domain scenarios, it can accurately identify 12 different behaviours of broilers, providing new insights for the development of intelligent equipment in livestock farming.