Housed pig identification and tracking for precision livestock farming

Abstract

Given the growing demand for pig products and improved welfare in the pig farming industry, Precision Livestock Farming (PLF) techniques are gaining more attention from farmers. Animal recognition and monitoring can provide major control in large-scale farms, reducing production costs, increasing welfare, and preventing the spread of diseases. However, animal re-identification is complex due to the scarce data available, the large similarity between the classes, and the variability of poses per class. To date, we cannot find any reliable solution in scientific literature or deployed in a real setting showing accurate stuff. This work uses current deep-learning models to define a complete pipeline to recognize pigs in different farm pens. For that, we present a complete pattern recognition pipeline comprising data collection, annotation, deep re-ID, and tracking to provide a reliable and accurate system for pig re-identification. Besides, we present two public datasets for segmentation (FaroPigSeg) and re-identification (FaroPigReID-33) that will contribute to the community of Precision Livestock Farming. Results show that a deep learning classification model (EfficientNetV2M) is good enough to reach a re-id accuracy of 90.44 %. Besides, the combination of a detector module together with a tracker, re-id, and voting modules can compute online predictions with an accuracy of 76.80 % in three challenging farming recording sessions.