Appearance-Based Gaze Tracking Through Supervised Machine Learning

Applications that use human gaze have become increasingly more popular in the domain of human-computer interfaces, and advances in eye gaze tracking technology over the past few decades have led to the development of promising gaze estimation techniques. In this paper, a low-cost, in-house video camera-based gaze tracking system was developed, trained and evaluated. Seminal gaze detection methods constrained the application space to indoor conditions, and in most cases techniques required intrusive hardware. More modern gaze detection techniques try to eliminate the use of any additional hardware to reduce monetary cost as well as undue burden to the user, all the while maintaining accuracy of detection. In this work, image acquisition was achieved using a low-cost USB web camera mounted at a fixed position on the viewing screen or laptop. In order to determine the point of gaze, the Viola Jones face detection algorithm is used to extract facial features from the image frame. The gaze is then calculated using image processing techniques to extract gaze features, namely related to the image position of the pupil. Thousands of images are classified and labeled to form an in-house database. A multi-class Support Vector Machine (SVM) was trained and tested on this data set to distinguish point of gaze from input face image. Cross validation was used to train the model. Confusion matrices, accuracy, precision, and recall are used to evaluate the performance of the classification model. Evaluation of the proposed appearance-based technique using two different kernel functions is also assessed in detail.