Neural Network-Driven Accuracy Enhancement for Wearable Eye-Tracking Devices Using Distance Information
Keywords: Head-mounted eye-tracking device, Neural Network-based Parallax Correction, Accuracy Improvement, Pupil Labs Invisible
Abstract: Eye-tracking devices are convenient for interpreting human behaviours and intentions, opening the way to contactless human-computer interaction for various application domains. Recent evolutions have enhanced them into wearable eye-tracking devices that opened the technology to the real world by allowing wearers to move freely and use them in regular indoor or outdoor activities. However, the gaze estimate from wearable devices remains more approximative than standard stationary eye-tracking devices due to their design constraints and a lack of interpretation of the three-dimensional scene of their wearer. This paper proposes to improve the gaze estimation accuracy of wearable eye-tracking devices using a framework that involves two neural networks, CorNN and CalNN. The CorNN corrects the bias induced by the distance between the observer and the gaze locations, primarily due to the parallax and lens distortion effects. While the CalNN is used to improve wearer-specific calibration. A robotic data collection system is implemented to automate training data acquisition for these networks. The proposed network has been demonstrated over a Pupil Labs Invisible eye-tracking device and tested on 11 wearers, showing improvement in the average gaze estimation accuracy on all wearers, especially at short-range reads.
Submission Type: Full Paper
Submission Number: 17
Loading