A computer vision-based system for cattle behavior monitoring and automated estrus detection

The profitability of cattle farming is predominantly dependent on effective reproductive outcomes, particularly high pregnancy rates, which are typically facilitated through artificial insemination. Nevertheless, the precise detection of estrus behavior presents a considerable challenge, significantly influencing the success of insemination procedures and the associated costs of farm operations. This thesis aims to revolutionize cattle reproduction management by implementing Computer Vision (CV) technology to monitor and interpret cow behavior for optimal insemination timing. Through remote camera surveillance and the development of specialized models, including cow detection, keypoint localization, and behavior classification, the project seeks to create a comprehensive system for real-time monitoring and decision support. The knowledge gained from this endeavor extends beyond technological advancements, offering insights into cattle behavior, ethical research practices, and the potential for transformative impacts on farming practices. Utilizing the YOLOv8 model, the research successfully developed a computer vision system that demonstrated high precision and recall rates in detecting and localizing cattle. The system was trained and validated on a dataset featuring diverse behaviors captured in farm environments. However, the dataset was noted as insufficient for optimal model training, highlighting the need for a more expansive collection of annotated data. A behavior classification model was also developed, achieving an overall accuracy of 64\%. This model incorporated features critical for understanding inter-cow relationships but was limited by the current capabilities in cow identification technology, representing a significant area for future research.The expert system built on these models supports farmers by reducing the reliance on manual observation and increasing breeding process efficiency. However, challenges in scaling these technologies to real-world applications were identified, particularly in terms of technological integration and operational adaptability. Future work will focus on expanding the dataset, enhancing model accuracy through advanced identification features, and refining integration with farm management systems. This thesis contributes to agricultural technology by demonstrating the potential of integrating computer vision and machine learning into effective farm management practices.