Livestock posture recognition using deep learning

Abstract

Calf posture recognition could be one of the required steps for a complete automated calf monitoring system, as sometimes the calf is required to be in a standing posture before being able to proceed to the next stage. To distinguish calf postures such as between standing or lying, machines require complicated frameworks, especially one that involves deep learning models. Previously, most of the works utilized video inputs rather than image inputs, which would make the model unnecessarily complicated compared to a conventional Convolutional Neural Network (CNN) model, which accepts image inputs. In this paper, to overcome all the problems mentioned earlier, two deep learning models with the exact same ResNet-50 based architecture have been built and trained on two different image datasets, respectively sourced from separate cameras placed at different angles to be compared and analyzed. The performance for both CNN models were 99.7% and 99.99% in accuracy, respectively, significantly better than the 92.61% accuracy of a similar work, and is adequate for a real-Time calf monitoring system.