Refined Pose Estimation for Square Markers Using Shape Fitting

We introduce an algorithm to refine the estimation of corners and pose of square fiducial markers, such as Arucos, with focus on mobile augmented reality applications. The idea is to reduce pixel jitter, which causes distracting artifacts such as “vibrating” objects, by exploiting information from the contour pixels of the detected markers. To achieve this, we develop a nonlinear least squares estimator that models a marker explicitly as a polygon and employs ideas from shape fitting. This provides not only a best-fitting estimate of the corners, but also a covariance matrix that can be used during further processing. We also implement a pose estimator that incorporates these covariance matrices and show how the effect of pixel jitter is greatly reduced in our approach, without increasing resource usage substantially in mobile devices.