Scholarlink

For individuals with motor or speech impairments, communication remains a critical challenge. Blink-to-Speak (BTS) language enables interaction through eye movements but requires intense focus and is prone to errors. This study proposes a computer vision-based alternative using video-oculography (VOG) signals to classify eye movements corresponding to eight BTS alphabets. A dataset of video recordings was processed via MediaPipe to extract facial landmarks and generate VOG signals. Features were extracted and used to train machine learning models. Random Forest (RF) achieved the best performance, with an average F1-score of 85.17%. The alphabets “Right” and “Turn” were most accurately classified (96.70% and 93.19% F1-score, respectively), while “Down” and “Wink” showed lower accuracy. This approach offers a non-invasive, scalable, and promising solution for communication systems supporting individuals with physical limitations.

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

VOG Multiclass Signal Classification for the Blink-to-Speak Language: Alphabets Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kevin B. Galindo-Pérez,
  • Valeria Fregoso-Jiménez,
  • José G. Sandoval-Ochoa,
  • Ricardo A. Salido-Ruiz,
  • Maria Cristina Padilla-Becerra,
  • Emilio Barajas-Gonzalez

摘要

For individuals with motor or speech impairments, communication remains a critical challenge. Blink-to-Speak (BTS) language enables interaction through eye movements but requires intense focus and is prone to errors. This study proposes a computer vision-based alternative using video-oculography (VOG) signals to classify eye movements corresponding to eight BTS alphabets. A dataset of video recordings was processed via MediaPipe to extract facial landmarks and generate VOG signals. Features were extracted and used to train machine learning models. Random Forest (RF) achieved the best performance, with an average F1-score of 85.17%. The alphabets “Right” and “Turn” were most accurately classified (96.70% and 93.19% F1-score, respectively), while “Down” and “Wink” showed lower accuracy. This approach offers a non-invasive, scalable, and promising solution for communication systems supporting individuals with physical limitations.