On Visual Gaze Tracking Based on a Single Low Cost Camera Evangelos Skodras, Vasileios G. Kanas, Nikolaos Fakotakis Department of Electrical and Computer Engineering, University of Patras, Patras, Greece Abstract Gaze tracking technologies provide an unconventional way of human-computer interaction, envisaged to advance practical applications and industrial products in a multitude of fields. The success of such systems depends on selecting the best calibration setup and image features that correspond to a person’s line of sight. The purpose of this study is to estimate eye gaze from a single, low cost web-cam, under natural lighting. Facial traits are extracted from the sensory data, from which distance vectors related to gaze are derived. Different experimental setups are studied, to evaluate the robustness of the proposed method with respect to various calibration setups, camera position and head movements. The use of new additional features improves the modeling of the subtle eye movements in the vertical direction, while a new calibration setup is proposed that further enhances the performance. The results demonstrate that the proposed framework is able to track gaze with good accuracy, consolidating the use of inexpensive equipment and techniques towards an ever-expanding range of gaze tracking applications. Keywords: Gaze estimation, Gaze tracker, Eye tracking, Human computer interaction, HCI 1. Introduction and Motivation The swift growth of functional sophistication in com- puting over the last decades has inevitably induced a growing interest in improving all aspects of interaction between humans and computers. This emerging field is gaining momentum for scientists across several differ- ent disciplines such as computer science, engineering, psychology and neuroscience. Along with speech, eye gaze comprises the most natural and comfortable means for human-computer interaction (HCI), giving rise to the ever-growing interest to develop systems that take advantage of gaze tracking technologies. Eye gaze is defined as the direction of a person’s line of sight, revealing a person’s focus of attention. It comprises a significant source of information about the cognitive and affective state of human beings, provid- ing implicit cues of intention and interest. As a con- trol input, in conjunction with the standard input meth- ods, gaze can greatly increase efficiency and usability. Gaze monitoring can be applied in a wide range of ap- plications [1], including non-glasses type 3D technolo- gies [2], monitoring of drivers attention and vigilance Email addresses: evskodras@upatras.gr (Evangelos Skodras), vaskanas@upatras.gr (Vasileios G. Kanas), fakotaki@upatras.gr (Nikolaos Fakotakis) [3, 4, 5], visual attention analysis (e.g. for marketing purposes [6]), interactive gaze-based interfaces for dis- abled people [7], diagnostic purposes [8, 9] and atten- tive HCI interfaces [10, 11]. Despite active research in the field, ubiquitous gaze tracking is beyond the grasp of the current systems. The vast majority of research in academia and industry is di- rected towards gaze tracking using active light sources, i.e. infrared (IR) illumination, achieving high accuracy rates [12, 13]. Hitherto, numerous commercial products making use of this well known approach are already on the market. However, active light approaches require dedicated hardware equipment which is usually high priced and the intrusiveness of which is controversial. Moreover, as they usually require a controlled environ- ment to prevent undesired reflections in the eyes, their applicability during day time is precluded. Other com- mon approaches employ 3D techniques [14] (using mul- tiple cameras or depth sensors) and wearable devices such as helmets or glasses [15], being cumbersome for the users. Universal gaze tracking from completely un- obtrusive, remotely located low-cost sensors (e.g. web- cams) still remains one of the most sought-after goals among researchers. Although web-cam based gaze trackers have so far demonstrated inferior performance compared to active Preprint submitted to Signal Processing: Image Communication April 9, 2015