Scholarlink

<p>Eye tracking is an essential tool for studying human mental activity and behavior. Appearance-based eye tracking has broad application prospects in actual unconstrained scenes, especially in the field of human-computer interaction. However, existing gaze estimation methods are less robust when the head pose changes. In order to improve the accuracy and real-time performance of eye tracking on the device, we propose a novel multi-input free head pose eye-tracking network model with the name FreeGazeNet, which uses binocular images and face images as inputs to estimate the gaze and the head posture, and designs a new loss function to adapt to model training. We use this method to evaluate the eye-tracking data sets MPIIFaceGaze and GazeCapture, and compared with the latest existing methods, the accuracy is improved by 3.4% and 10.0%, respectively. In addition, we also propose a simple post-processing method for individual gaze calibration to obtain more accurate estimates for different users. The experimental results prove that our proposed method can obtain higher accuracy in gaze estimation and has a greater value in practical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Free-Head Gaze Estimation with Deep Learning

  • 期刊论文

  • 发表日期：

    2025年4月24日

查看全文
  • Feng Duan,
  • Zhuochao Song,
  • Xuyi Chen,
  • Shifan Yin,
  • Zhe Sun,
  • Bing Liang,
  • Yisha Guo

摘要

Eye tracking is an essential tool for studying human mental activity and behavior. Appearance-based eye tracking has broad application prospects in actual unconstrained scenes, especially in the field of human-computer interaction. However, existing gaze estimation methods are less robust when the head pose changes. In order to improve the accuracy and real-time performance of eye tracking on the device, we propose a novel multi-input free head pose eye-tracking network model with the name FreeGazeNet, which uses binocular images and face images as inputs to estimate the gaze and the head posture, and designs a new loss function to adapt to model training. We use this method to evaluate the eye-tracking data sets MPIIFaceGaze and GazeCapture, and compared with the latest existing methods, the accuracy is improved by 3.4% and 10.0%, respectively. In addition, we also propose a simple post-processing method for individual gaze calibration to obtain more accurate estimates for different users. The experimental results prove that our proposed method can obtain higher accuracy in gaze estimation and has a greater value in practical applications.