Developed primarily for a research setting, a computer vision system was created to monitor sows and piglets housed in farrowing stalls. Above each farrowing stall a Microsoft Kinect V2 was mounted to capture digital and depth images of the animals below. Images were taken once every five seconds for the entire duration of the five-week farrowing and lactation cycle. Specialized processing algorithms were developed to interpret the information captured in the images and identify the sow’s posture as sitting, standing, kneeling or lying. When labeled as lying the orientation of the udder was labeled. If occupying one of the other postures, the sow was labeled as eating, drinking or other behavior. Program labels were compared to the ground truth as determined by human observers to determine label accuracies. This system labeled sow posture with an accuracy greater than 99%, and greater than 95% accuracy for all behavior labels.
This system can be used to evaluate the effects of changing features of the farrowing environment on sow and piglet behavior. For example, with this system researchers can monitor the amount of time individual sows spend lying, orientation of the udder and if the sow is feeding and drinking. Sow behavior patterns can be compared among experimental treatments to determine if a particular treatment (crate layout, supplemental heat source type and placement, etc.) results in more desirable sow behaviors (e.g., less frequent postural changes). Coupled with production records, this information can be used to evaluate various farrowing stall designs for productivity and sow health to produce recommendations for farrowing stall designs.
An alternative approach is hardware-based to determine the sow’s posture. This method detects sow posture (standing, lying or other) by utilizing feedback from three depth sensors. These three sensors are installed above the sow on the farrowing crate and spaced equally along the length of the stall. While it cannot monitor a sow’s eating behavior, it does have the advantage of determining the sow’s posture in real-time for lower development and computational costs. Through the initial testing and development phases the system accuracy is averaging at 90% to a machine vison system for posture detection. Applying this type of system to a production system could offer producers the ability to detect the onset of farrowing and dangerous behaviors (high frequency of posture changes) during lactation.
Lire la suite: www.nationalhogfarmer.com
