V.DIG.0028 - Computer Vision for Individual Animal Identification

Summary

The focus of this research project was to determine whether a fully computer vision-based system could accurately identify individual cattle with sufficient reliability to supplement traditional identification technologies, such as RFID tags. A software-based solution of this kind offers a non-intrusive approach to cattle identification that can be automated without manual handling or data collection, while also integrating readily with other vision-based intelligence and supply chain tracking systems.

Objectives

The following objectives were achieved at the conclusion of this project, through the development, evaluation and delivery of a functional prototype recognition system:

•    Review, analyse, and curate existing cattle image and video datasets to assess data quality and available visual features, while establishing a benchmark validation environment and evaluation protocol for recognition testing.
•    Train and validate prototype computer vision recognition models capable of generating discriminative cattle identity embeddings using multiple visual cues, to assess the feasibility of AI-based cattle identification under both controlled and practical conditions.
•    Design and implement a modular ensemble-based recognition system that combines multiple visual recognition features, to improve identification accuracy and robustness.
•    Conduct a comprehensive performance analysis of the overall recognition system to evaluate operational feasibility, scalability, computational requirements, and deployment considerations.

Key findings

The prototype achieved 100% top-1 identification accuracy on the controlled Cattle Weight V.TEC.1731 dataset, demonstrating strong effectiveness under ideal imaging conditions. On the more complex P.PSH.1424 dataset, designed to reflect real-world farm environments, the system achieved 62.5% top-1 accuracy, confirming the feasibility of computer vision-based cattle identification while also highlighting the challenges introduced by operational variability. Importantly, the results indicate clear and achievable opportunities for further performance improvements through additional feature integration and model refinement with continued development effort.

Benefits to industry

The research demonstrates the potential for non-intrusive, vision-based cattle identification systems to improve livestock traceability and automation within the meat supply chain. Such systems could reduce reliance on RFID technologies, lower handling stress for animals, and support integration with other automated computer vision applications such as behaviour monitoring and weight estimation. The expandable prototype also provides a foundation for future commercial and research development.

MLA action

MLA continues to invest in the advancement of computer imaging technologies for the advancement of livestock productivity and wellbeing.

Future research

Future research should focus on improving robustness in real-world farm environments through the integration of additional visual features, more advanced data fusion strategies, and enhanced model refinement techniques. In addition, while this project was limited to frame-based image analysis, future work should extend to full video-based interpretation, incorporating temporal information through tracking and sequence-level modelling to better reflect operational deployment conditions. Further testing across varied lighting conditions, cattle breeds, and camera perspectives is also recommended to improve generalisability and operational reliability. Based on the findings of this project, achieving deployment-ready accuracy levels of approximately 90% appears feasible with continued development and optimisation.