P.PSH.1279-Proof of Concept trial for retinal scanning of cattle for identification

Summary

Australia is recognised as a world leader in livestock identification and traceability. Systems like Australia’s National Livestock Identification System (NLIS) are recognised and respected around the globe as ‘best practice’.

Biometrics provide an opportunity to identify farm animals permanently and securely without requiring individual tags. This project undertook a proof-of-concept pilot using retinal scanning for individual animal identification, which has been developed for the security industry and is now being made available for animal applications.

This was an initial trial to demonstrate that the technology can be used on cattle, as a first step in assessing its application to the livestock sector, identify any challenges to its use, and to identify next steps in commercialising it in the livestock industries.

Objectives

• Determine accuracy of iTRAK retinal matching algorithm in correctly identifying individual cows under commercial environment.
• Explore practicality of retinal image scanning for animal identification on commercial farms.

Key findings

The current variant of the iTRAK classification algorithm had modest sensitivity at eye level. The sensitivity of left eye scans was 82.4% (95 %CI = 71.5-90.0%; 61 of 74 scans identified the cow) and for the right eye scans was 75.7% (95% CI = 64.1-84.6%; 56 of 74 scans identified the cow).

Applying a testing-in-parallel approach (i.e. at least one of the left or right scans confirming identification is required to identify the animal) the sensitivity was 93.2% (95% CI = 84.3-97.5%; 69 out of 74 scans identified the cow).

There are operational challenges to use of this technology on farm animals. Ambient light causes excessive pupillary constriction to allow suitable retinal image capture using the iTRAK camera under typical daylight conditions. The current hardware is difficult to use whilst handling the cows’ head (hard to see the viewfinder) and the presence of eye disease (cataracts and pink eye) can prevent suitable image capture from individual cows.

The algorithm may be prone to overfitting. This can be improved by construction of a much larger trial set of data – obtained from multiple animals, multiple farms and multiple operators such that a training and testing dataset and cross-validation techniques can refine the algorithm.

Benefits to industry

Biometric-based animal identification provides many advantages over existing tag-base systems including cost, reliability, robustness and tamper resistance. However, there are challenges to recording biometric data in large farm animals and logistical restrictions may exists - such as the necessity to access on-line databases when assuring the provenance of an individual animal.

Future research

Refinements to the algorithm are essential. More training data is essential to allow the algorithm to improve in accuracy and robustness.

Adjustments to hardware and software and more robust data transfer systems may be required to provide commercial usefulness for the system. Cameras that can capture the whole of the retina through slit-like pupils and which can operate in daylight conditions are also recommended.