Camgaze.js: Browser-Based Eye Tracking with Standard Cameras

As it has always been hardware-dependent, eye tracking has been difficult and costly to use, needing a dedicated piece of hardware or a particular software bundle that is designed for a specific type of hardware. This paper introduces Camgaze.js, a JavaScript library for gaze estimation in a web browser without dedicated hardware or native software. The client-side pipeline includes web-camera video acquisition (via WebRTC), eye localization (Viola-Jones algorithm), pupil detection (thresholding and connected-component analysis of a binary image), and gaze estimation (linear calibration of the gaze position in the camera image). Video frames are acquired, calibrated, and processed entirely within the browser window so that they are not transmitted or stored. (However, host-page scripts can access the given gaze position values exposed by the library.) For the system evaluation, 5 sessions of accuracy estimation and 5 sessions of performance measurement were conducted with a host laptop with a built-in camera. The experiments showed an average radial gaze error of 2.1 inches with a standard deviation of 0.1 inches for 10 tracked targets per session, and a usable sampling rate of 17.5 fps with a standard deviation of 4.5 fps across sessions. These preliminary assessments show promising potential for lightweight, accessible eye tracking that provides coarse gaze estimates, but also confirm limitations such as accuracy, lighting conditions, head pose, and the scale of evaluation.