Sow posture detection for determining piglet crushing through a camera system

Abstract

Piglet crushing by sows is a leading cause of pre-weaning mortality on commercial pig farms. While improved management can mitigate this, resource limitations often hinder timely intervention. This article proposes an automated warning system that analyzes sow posture from surveillance footage in real-time to predict and prevent crushing events. We developed a posture detection model using You Only Look Once (YOLO)v8, trained on 422 real-world instances, which achieved a mean Average Precision (mAP@50) of 0.976. Our analysis revealed a significant increase in the frequency of sow postural changes on the day of a crushing event (8.23) and the day prior (7.49), compared to non-crushing days (5.52). Leveraging this insight, we designed a warning system using a threshold-based voting algorithm that analyzes posture changes over a 60-min window. The system's performance was evaluated at two levels. For instance-based warnings (on a 60-min basis), it achieved a sensitivity of 62.50% and a specificity of 60.25%. When aggregated to a daily basis, the performance improved to a sensitivity of 71.42% and a specificity of 84.61%, respectively. Our results indicate that sow postural change frequency is a promising indicator for developing early warning systems, empowering farmers to take preventative action and reduce piglet losses.