Scholarlink

<p>In 2023, the World Health Organization (WHO) conducted a global survey revealing that 16% of the world’s population lives with some form of functional limitation. The majority of these individuals reside in developing countries, where access to assistive technologies is limited due to economic and structural barriers. In this context, the development of affordable and accessible solutions is crucial to promoting autonomy and inclusion. This study proposes a facial movement-based wheelchair control system designed for individuals with tetraplegia. The system leverages real-time computer vision techniques and the YOLOv8 model to detect specific facial features - with a focus on nose position and orientation - converting these into directional commands (forward, backward, left, and right). The framework is structured into three main stages: facial detection, feature extraction, and command inference. To classify directional commands, we evaluated two machine learning methods: PCA and XGBoost. The confusion matrices demonstrated that XGBoost achieved superior accuracy across all classes, particularly in the "forward" direction (89%), outperforming PCA (78%). Interpretability analysis using a SHAP summary plot showed that the most influential features for decision-making are delta_x and delta_y, corresponding to the horizontal and vertical displacements of the nose. These features are directly linked to rotational and linear movement commands. Angular attributes such as angle_yaw and angle_pitch, which represent head rotation in the horizontal and vertical planes, further support directional inference, especially for users with limited motion amplitude. Conversely, features like angle_roll and stability_score demonstrated low relevance, as they are not directly associated with mapped commands in the current system. Overall, the performance metrics - 96.5% precision, 95.8% recall, and 95.9% mean Average Precision (mAP) - confirm the robustness of the proposed approach. The integration of advanced visual processing with interpretable models results in a responsive and intuitive control interface, effectively supporting users with severe motor impairments and enhancing their independence and quality of life.</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

Using deep learning applied in computer vision for the inclusion of individuals with tetraplegia through assistive robotics via facial features

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Fernando Rodrigues Trindade Ferreira,
  • Loena Marins do Couto,
  • Camila Martins Saporetti

摘要

In 2023, the World Health Organization (WHO) conducted a global survey revealing that 16% of the world’s population lives with some form of functional limitation. The majority of these individuals reside in developing countries, where access to assistive technologies is limited due to economic and structural barriers. In this context, the development of affordable and accessible solutions is crucial to promoting autonomy and inclusion. This study proposes a facial movement-based wheelchair control system designed for individuals with tetraplegia. The system leverages real-time computer vision techniques and the YOLOv8 model to detect specific facial features - with a focus on nose position and orientation - converting these into directional commands (forward, backward, left, and right). The framework is structured into three main stages: facial detection, feature extraction, and command inference. To classify directional commands, we evaluated two machine learning methods: PCA and XGBoost. The confusion matrices demonstrated that XGBoost achieved superior accuracy across all classes, particularly in the "forward" direction (89%), outperforming PCA (78%). Interpretability analysis using a SHAP summary plot showed that the most influential features for decision-making are delta_x and delta_y, corresponding to the horizontal and vertical displacements of the nose. These features are directly linked to rotational and linear movement commands. Angular attributes such as angle_yaw and angle_pitch, which represent head rotation in the horizontal and vertical planes, further support directional inference, especially for users with limited motion amplitude. Conversely, features like angle_roll and stability_score demonstrated low relevance, as they are not directly associated with mapped commands in the current system. Overall, the performance metrics - 96.5% precision, 95.8% recall, and 95.9% mean Average Precision (mAP) - confirm the robustness of the proposed approach. The integration of advanced visual processing with interpretable models results in a responsive and intuitive control interface, effectively supporting users with severe motor impairments and enhancing their independence and quality of life.