Deep Learning-Based Piglet Tracking Algorithm for Automated Crushing Detection on Pig Farms

Abstract In modern pig farming, piglets and sows are housed together in farrowing pens during the lactation period. However, piglets face the risk of being trapped and suffocated by the sow, a major cause of piglet mortality that significantly impacts farm productivity. To address this issue, various methods have been developed to detect piglet crushing using either acoustic or image data. However, these approaches often fail when the piglet is completely obstructed by the sow, rendering it hidden to both audio and image detection. In this paper, we propose a two-stage approach, consisting of a Hidden Trapping Prediction Algorithm (HPA) and a Crushing Decision Algorithm (CDA). The HPA tracks the total number of piglets, monitors those that are unseen, and detects changes in their count, allowing us to predict which unseen piglets are likely to be trapped. The CDA uses a YOLO model to track newly detected piglets, estimating their movements and identifying objects that remain stationary for a certain period as likely to be crushed. We also developed the Trapping Prediction Algorithm (TPA), which combines an image-based trapping detection model with the HPA. This model assesses the number of objects per frame and analyzes the movement of new objects. To evaluate our scheme, we collected the video footage from five litters housed in loose farrowing pens, capturing both trapping and crushing events. Our performance evaluation confirmed that the proposed scheme efficiently tracks piglets and predicts hidden trapping events and crush events with a mean absolute error (MAE) of 1 and 0.6 respectively. Furthermore, our scheme achieved high accuracy in detecting piglet trapping events, with a R2 value of 0.49, surpassing existing image-based models.