Identificação, rastreamento e contagem de suínos em tempo real através de imagens

Abstract

This study proposes the development of an embedded system for real-time identification, tracking, and automatic counting of pigs, aiming to minimize errors during the loading process—a critical stage in the pig production chain. Using the NVIDIA Jetson Nano platform and the YOLOv12 detection algorithm across its model variations, combined with the SORT and ByteTrack trackers, a comprehensive comparative analysis was conducted focusing on accuracy, inference time, and identifier stability. The system was evaluated in real-world production conditions, with videos captured at a commercial farm, totaling six distinct scenarios and over one thousand annotated images. Results showed that the YOLOv12-nano model, paired with the ByteTrack tracker, achieved the best performance, delivering high counting accuracy and robustness against challenges such as occlusions, animal overlap, and lighting variations. This research contributes to the modernization of Brazilian pig farming by offering a low-cost, efficient, and practical solution for automating animal flow control, positively impacting livestock management and production logistics.