Development of a Fast and Robust Gaze Tracking System for Game Applications

In this study, a novel eye tracking system using a visual camera is developed to extract human’s gaze, and it can be used in modern game machines to bring new and innovative interactive experience to players. Central to the components of the system, is a robust iris-center and eye-corner detection algorithm basing on it the gaze is continuously and adaptively extracted. Evaluation tests were applied to nine people to evaluate the accuracy of the system and the results were 2.50 degrees (view angle) in horizontal direction and 3.07 degrees in vertical direction.

💡 Research Summary

The paper presents a gaze‑tracking system designed for modern game applications that relies solely on a conventional visual‑light camera rather than infrared illumination. The authors argue that infrared‑based video‑oculography, while accurate, can cause eye dryness, lens burning, and retinal injury, motivating a low‑cost, user‑friendly alternative. The core of the system consists of three processing stages: (1) iris‑center detection, (2) eye‑corner detection, and (3) gaze estimation using an eyeball geometric model.

Iris‑center detection begins with down‑sampling a 640 × 480 eye image to 80 × 60, applying a dynamic iso‑data threshold, and extracting the two largest connected components (eyebrow and eye). A sliding window whose width is 15 % of the eye region scans horizontally; the position with maximal gray‑value sum is taken as an initial iris estimate. The algorithm then returns to full resolution, refines the threshold, and extracts sample points along the iris‑sclera border using a zig‑zag scanning pattern that reduces noise and computational load. These points are fed into a double‑circle fitting procedure: an algebraic least‑squares fit is performed twice, discarding outliers after the first fit, to obtain a robust iris center.

Eye‑corner detection uses a variance projection function (VPF) in the vertical direction. After defining a region around the iris based on its radius, the VPF and its derivative are computed; the peak of the derivative indicates the vertical location of the eye corner. A Sobel edge detector finds the eyelid, and the horizontal coordinate is taken as the intersection of the edge with the VPF peak. The eye corner serves as a reference point that compensates for head translation without requiring external markers or lengthy calibration.

Gaze estimation combines the iris‑center, eye‑corner, and an eyeball model. The vector from eye corner to iris center is measured; with assumed eyeball radius (12–13 mm) and screen‑to‑eye distance (≈650 mm), the model projects the spherical eye onto the screen plane. Two calibration points are used to compute a linear scaling factor k, which corrects systematic errors in the mapping from measured displacements (Δ₁, Δ₂) to screen coordinates (g₁, g₂).

The hardware prototype consists of a Texas Instruments TMS320C6416 DSP board (1 GHz) and a Sony camera, achieving 30 fps capture. The algorithm can be applied ten times per frame, implying a maximum sampling rate of 300 Hz.

Experimental evaluation involved nine participants (six with bare eyes, three wearing glasses). After a two‑point calibration, subjects were asked to fixate on seven target points while auditory cues guided them. The average horizontal error was 28 mm (≈2.50° visual angle) and the average vertical error was 35 mm (≈3.07°). The best case error was 21 mm (≈1.85°). Errors were larger vertically, likely due to eyelid occlusion. Failure cases occurred when a large portion of the iris was hidden (e.g., by hair covering the eyebrow), leading to insufficient sample points for circle fitting.

To demonstrate applicability, the authors implemented a simple game “Hammer Heads” where players shoot monsters by gazing at them, and a survey machine where users answer Yes/No questions by looking at on‑screen buttons. Users reported a novel and enjoyable interaction, though occasional detection failures required re‑tries.

In conclusion, the study delivers a fast, robust, and inexpensive gaze‑tracking solution suitable for game interaction. It achieves sub‑3° accuracy and high frame‑rate processing without infrared hardware or wearable markers. The authors acknowledge current limitations in precision and plan to improve distance estimation and calibration methods, potentially integrating more advanced image‑processing or machine‑learning techniques to further enhance reliability and accuracy.