Scholarlink

<p>Meal-assisted robots (MARs) can significantly aid individuals with limited upper-limb mobility to eat independently. To develop an autonomous feeding system for MARs, this paper presents a novel approach that integrates the You Only Look Once model with a binocular camera for precise mouth recognition and positioning. In our mouth recognition experiments, the proposed method achieves an optimal balance between accuracy and speed, attaining a test set accuracy of 97.5% and an average detection frame rate of 24 fps, meeting real-time detection requirements. For mouth localization, we utilize the depth map generated by the binocular camera in conjunction with the output of the optimal recognition model. To ensure feeding safety, we evaluate the recognition range under various accuracy thresholds, finding that the optimal threshold range is 0.5 to 0.8, with the best camera distance positioned between 20&#xa0;cm and 80&#xa0;cm from the mouth. Additionally, our positioning system achieves an average positioning error rate of approximately 0.33%, ensuring reliable operation. This study not only establishes a clear relationship between mouth detection and spatial target positioning but also provides substantial support for the development of autonomous feeding systems for MARs. The code and data declaration are available at <a href="https://github.com/beibeidemei9420/mouth_location">https://github.com/beibeidemei9420/mouth_location</a>.</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

High-accuracy real-time mouth recognition and 3D positioning for autonomous feeding robots using YOLO and binocular vision

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • Xin Li,
  • Kai Zhang,
  • Qing Yuan,
  • Xiafeng Shen,
  • Rui Zhu,
  • Yue Kai

摘要

Meal-assisted robots (MARs) can significantly aid individuals with limited upper-limb mobility to eat independently. To develop an autonomous feeding system for MARs, this paper presents a novel approach that integrates the You Only Look Once model with a binocular camera for precise mouth recognition and positioning. In our mouth recognition experiments, the proposed method achieves an optimal balance between accuracy and speed, attaining a test set accuracy of 97.5% and an average detection frame rate of 24 fps, meeting real-time detection requirements. For mouth localization, we utilize the depth map generated by the binocular camera in conjunction with the output of the optimal recognition model. To ensure feeding safety, we evaluate the recognition range under various accuracy thresholds, finding that the optimal threshold range is 0.5 to 0.8, with the best camera distance positioned between 20 cm and 80 cm from the mouth. Additionally, our positioning system achieves an average positioning error rate of approximately 0.33%, ensuring reliable operation. This study not only establishes a clear relationship between mouth detection and spatial target positioning but also provides substantial support for the development of autonomous feeding systems for MARs. The code and data declaration are available at https://github.com/beibeidemei9420/mouth_location.