Application of deep learning to automated inspections of pig farms by unmanned vehicles

We propose a smart inspection system for livestock barn environments with an IP camera mounted on an unmanned vehicle to perform automated inspections of narrow pens for the monitoring of sow behavior. The unmanned vehicle uses ultra-wideband (UWB) technology and proportional-integral-derivative (PID) control to maintain stable linear navigation in the barn. The YOLOv4 deep learning architecture was applied to analyze images of sow behavior and spray-painted pen numbers on the feeding troughs for image recognition. Our database contained a total of 838 images and four labels: (1) Num, (2) Standing, (3) Lying_down, and (4) Empty. Testing of video samples resulted in 100% detection accuracy. The study results show that combining UWB positioning, PID control, and YOLOv4 image recognition can effectively achieve automated inspections and highly reliable monitoring of sow behavior in barns.