A machine learning-based positioning method for poultry in cage environments

Abstract

Individuals or groups of animals exhibit different activities that characterize domain behavior. Rapid and accurate localization of poultry in small and complex cage environments helps analyze the poultry domain behavior. This study proposes a machine-learning-based method for locating poultry in small and complex cage environments. Here, the characteristics of ultra-high frequency-radio frequency identification devices were determined, received signal strength indicator values were collected, and the tag-coordinate regression problem was converted into a multi-area classification problem. Different models were used to predict the target position. The results revealed that the neural network model yielded the best prediction, locating the target within a 40 cm × 40 cm area with 88.74% accuracy or within a 30 cm × 30 cm area with 76.81% accuracy, with average errors of 7.61 cm and 7.97 cm, respectively. Finally, experiments with live chickens were performed, and the results were verified using synchronized video, obtaining a Pearson correlation coefficient exceeding 0.909. This study presents a feasible method for target localization in small and complex cage environments, providing valuable modal information for multimodal learning.