Sow Posture Analysis and Estrus Prediction Using Closed-Circuit Television Cameras

Abstract

Sow estrus detection is one of the most critical tasks for improving the production performance of pig farms. However, accurately determining the onset of estrus is challenging because it is time consuming to check each sow and their performance, particularly during specific working hours. Moreover, estrus determination criteria are not standardized, as managers rely on their individual experiences. In this study, we proposed a method for predicting sow estrus using deep learning techniques. To detect sows and classify their postures, we used a lightweight deep-learning-based object detection model, You Only Look Once version 5 (YOLOv5). We trained one of the prediction models, Bidirectional Long Short-Term Memory (Bi-LSTM), which is a supervised learning model, using the time series data composed of a combination of each posture and holding time. By setting the ground truth as data from 24 h before the manager's estrus determination, we achieved an estrus prediction accuracy of 86 %. This study demonstrates the potential of using closed-circuit television (CCTV) footage to predict sow estrus, and the proposed method can contribute to reducing the labor required for sow estrus checks.