Methods for neural network detection of farm animals in dense dynamic groups on images

The development of non-invasive methods for monitoring the condition of farm animals is now a burning problem. The world is developing technologies for video surveillance of animals with subsequent image processing using neural networks. The purpose of this study is to develop methods for the detection (selection of individuals) of farm animals in images using pigs as an example. The main task is to perform the detection of “faces” of pigs in dense groups. To solve the task, a set of photographs of pigs from open sources was created, promising neural network architectures Faster R-CNN and YOLOv5 were selected, fine-tuning and training of neural networks were performed. The use of the YOLOv5 network enabled the detection accuracy mAP = 94.05%, which is significantly higher than the accuracy shown in similar works. This work is the first in an upcoming series of studies aimed at creating a software and hardware complex for automatic animal health monitoring on farms.