Home » Eye Tracking » Eye Tracking: Uncertainty in Iris Localization

Eye Tracking: Uncertainty in Iris Localization

Eye Tracking: Uncertainty in Iris Localization In a study coming from Cavendish Laboratory, Cambridge, UK, researchers have come up with a Bayesian solution to the problem of locating irises in an image known to contain a face. We’ve written a few articles about new security cameras that can provide iris recognition from afar or for unwilling or non-compliant suspects in a security check, so adding some detail on the processes behind some of these ideas could be helpful to our readers.

Locating irises has been a problem for many eye tracking devices, and solving this task is really one step in solving a larger problem. This study aligned images for face recognition, head pose estimation, and gaze tracking. They set out to use iris location for gaze tracking with low cost hardware like a standard web camera, so the system presented in the paper doesn’t assume access to infrared lighting or high resolution of the eye. It’s made to run in real time.

They’ve essentially approached the problem using a Bayesian framework to design a system that provides estimates of iris location. As for uncertainty in measurement, they’ve incorporated that too. Using their method, they say, enables computing a less certain estimate with little computation, and a more certain estimate with more computation. Uncertainty information is helpful when integrating a localizer into a larger system like head pose estimation or gaze tracking, which can use the certainty as a measure of confidence when it comes to iris estimates.
In addition, this approach can be used to locate other visual features as it learns the structure of images around the feature of interest rather than assuming any specific structure.

The authors write that in recent years there has been much work on increasing the accuracy of medium to low resolution iris localizers, but little work on providing estimates of uncertainty, something the researchers set out to do. The majority of systems provide point estimates, so some systems could theoretically be modified to provide uncertainty information by computing a particular approximation around the optimum of their energy functions.

In the end, the paper presents a probabilistic method for locating irises in images containing faces – one that shows localization accuracy comparable to the state of the art when allowing approximately one second per iris localization.

Take a look at the paper to see the details of the study – it’s something that could certainly be helpful for low-cost eye tracking endeavors.

Bayesian Iris Localization with Linear Filters

Eye Tracking: Uncertainty in Iris Localization

Categories