돼지의 행동 분류를 위한 딥러닝 기반 Instance Segmentation 모델

Abstract

In pig-breeding livestock farms, increasing yearly sow productivity and reducing piglet mortality are very important for farmers because they are directly related to profitability. To meet these requirements of the livestock industry, smart livestock technology is being developed and propagated. Recently, many studies have been conducted to obtain information on the health and physiological status of livestock by applying image processing- and artificial intelligence-based technology. In particular, because changes in pig behavior patterns can provide considerable information about their health and physiological status, this study proposes a method to classify pig behavior patterns from images of a pig kennel by applying a deep learning-based instance segmentation model. For this purpose, we first constructed a Pig Motion dataset by building a robot operation system (ROS)-based data collection system that could synchronize and collect voice and climate information along with pig motion images in time. This dataset was organized by classifying four types of pig behavior, namely, lying, standing, sitting, and eating, into motion classes. By learning this dataset using the instance segmentation model, an object is extracted from the pig’s motion characteristic information to create a model that divides and classifies it. We propose a behavioral pattern classification model that can be used to estimate the health and physiological state of pigs by classifying their behavioral patterns from streaming videos and accumulated data and statisticalizing them.