overall objective of this research project was to identify available technologies that can rapidly, and accurately, assess pig body weight at market. In order to complete the objectives, the research team; validated the accuracy of multiple commercially-available products designed to capture pig weights, developed a unique methodology to assess pig body weight using stereo vision cameras and advanced image processing, and determined if any of the available technologies are more effective at assessing pig body weight compared to visual evaluation by animal caretakers. The available pig weight assessment methods included in this study were: visual observation by an animal caretaker, a walk-across platform scale (CIMA; Correggio, Italy) that captures weights of pigs in real-time as they move freely over the scale, PigVision cameras (Asimetrix Inc.; Durham, NC) mounted directly above pens of pigs and a novel, handheld, portable RGB & stereo vision system (developed by the researchers) which uses images from various angles. In the first study, a trained individual estimated the market weight at two sites which included approximately 1,000 pigs. In addition, both RGB and depth data were collected from these pigs using the novel handheld device. A 16-week study was then conducted to determine the accuracy of PigVision cameras over time from placement to market. Cameras were mounted above 12 pens and each pen contained 8 pigs. Weights were validated every two weeks using a calibrated pen scale. A final study using 91 pigs at market weight compared the accuracy of visual evaluation, the walk-across scale, PigVision cameras and the stereo vision system developed by the researchers.

In the first study, a trained individual selected pigs estimated to be market weight at two sites. At site one, visual evaluation of 468 pigs had an accuracy of 84.4%, site two with visual observation of 522 pigs and an 82.5% accuracy. Accuracy was measured by whether the pig was marked correctly in the market weight range after the pig was weighed on a calibrated scale. When testing the accuracy of PigVision cameras over 16 weeks it was found that this system was less accurate for pigs weighing approximately 70 pounds (87.7%) than pigs weighing approximately 275 pounds (96.6%). A comparison of all methods at marketing of 91 pigs it was found that the walk-across scale provided the most accurate weights (98.2%), visual observation was the least accurate (88.2%), while PigVision provided a 96% accuracy. Using the RGB and depth data collected using an Intel RealSense camera from these pigs, our proposed method created pig weight estimation models using various data. Our best estimation of pig weight was within 20 pounds of actual weight but had a R^2 value of only 0.4167. This prediction was not as accurate when compared to the linear regression model using heart girth to predict weight with a R^2 value of 0.8621.

Human observation is the chosen method in many operations today yet offers the lowest accuracy. The walk-across scale is easy to operate but requires tactical animal movement. PigVision is the least arduous option, provides constant data, but does require some maintenance. The research introduced a new method to predict pig weight at market using angle-agnostic measurements. The research also improved rapid pig body weight estimation methods by combining both deep learning outputs and handcrafted features to provide economic benefits to swine producers through the development of a novel, handheld, mobile technology. Further work characterizing pigs of a larger weight range would improve the accuracy of the prediction models developed by these researchers.
This work was funded by the National Pork Checkoff. If you have further questions about this research please contact Jonathan Holt, PhD, Extension Swine Specialist at North Carolina State University. [email protected] or 919.515.4805.

Key Findings

  • Visual evaluation of pigs was the least accurate method to determine market weight of pigs
  • A commercially-available, mounted camera system was over 95% accurate at detecting market weights of pigs
  • Development of an angle-agnostic, computer vision method estimated pig market weights within 20 pounds of actual weight