Real-time appearance-based gaze tracking.
Abstract
Gaze tracking technology is widely used in Human Computer Interaction applications
such as in interfaces for assisting people with disabilities and for driver attention monitoring.
However, commercially available gaze trackers are expensive and their performance
deteriorates if the user is not positioned in front of the camera and facing it. Also, head
motion or being far from the device degrades their accuracy.
This thesis focuses on the development of real-time time appearance based gaze
tracking algorithms using low cost devices, such as a webcam or Kinect. The proposed
algorithms are developed by considering accuracy, robustness to head pose variation and
the ability to generalise to different persons. In order to deal with head pose variation, we
propose to estimate the head pose and then compensate for the appearance change and
the bias to a gaze estimator that it introduces. Head pose is estimated by a novel method
that utilizes tensor-based regressors at the leaf nodes of a random forest. For a baseline
gaze estimator we use an SVM-based appearance-based regressor. For compensating
the appearance variation introduced by the head pose, we use a geometric model, and
for compensating for the bias we use a regression function that has been trained on a
training set. Our methods are evaluated on publicly available datasets
Authors
Kaymak, SertanCollections
- Theses [3706]