Aplicação de técnicas de visão computacional para análise comportamental de animais em confinamento

Abstract

This study aims to evaluate and apply computer vision techniques for the behavioral analysis of confined cattle, focusing on the Holstein breed, to improve animal welfare. A custom database was created using day and night videos, supplemented with images from the CBVD dataset. The images were divided into frames and manually annotated with the Roboflow system, undergoing preprocessing to ensure quality and consistency. YOLOv8 and YOLOv9 versions were explored and compared, with the latter showing superior performance in Precision, Recall, and F1 score metrics. The analysis was conducted in the Google Colab Pro environment, with detailed specifications. The study concludes that YOLOv9 is more suitable for the proposed task, emphasizing the importance of spatiotemporal data for future improvements.