Yao Ching-Bang, Lu Cheng-Tai
Department of Information Management, Chinese Culture University, Taipei 11114, Taiwan.
Sensors (Basel). 2024 Nov 22;24(23):7448. doi: 10.3390/s24237448.
As many countries face rapid population aging, the supply of manpower for caregiving falls far short of the increasing demand for care. Therefore, if the care system can continuously recognize and track the care recipient and, at the first sign of a fall, promptly analyze the image to accurately assess the circumstances of the fall, it would be highly critical. This study integrates the mobility of drones in conjunction with the Dlib HOG algorithm and intelligent fall posture analysis, aiming to achieve real-time tracking of care recipients. Additionally, the study improves and enhances the real-time multi-person action analysis feature of OpenPose to enhance its analytical capabilities for various fall scenarios, enabling accurate analysis of the approximate real-time situation when a care recipient falls. In the experimental results, the system's identification accuracy for four fall directions is higher than that of Google Teachable Machine's Pose Project training model. Particularly, there is the significant improvement in identifying backward falls, with the identification accuracy increasing from 70.35% to 95%. Furthermore, the identification accuracy for forward and leftward falls also increases by nearly 14%. Therefore, the experimental results demonstrate that the improved identification accuracy for the four fall directions in different scenarios exceeds 95%.
随着许多国家面临人口快速老龄化，护理人力的供应远远无法满足日益增长的护理需求。因此，如果护理系统能够持续识别和跟踪护理对象，并在其跌倒的第一时间迅速分析图像以准确评估跌倒情况，将至关重要。本研究将无人机的机动性与Dlib HOG算法以及智能跌倒姿势分析相结合，旨在实现对护理对象的实时跟踪。此外，该研究改进并增强了OpenPose的实时多人动作分析功能，以提高其对各种跌倒场景的分析能力，从而能够在护理对象跌倒时准确分析近似实时的情况。在实验结果中，该系统对四个跌倒方向的识别准确率高于谷歌可教机器的姿势项目训练模型。特别是，向后跌倒的识别有显著提高，识别准确率从70.35%提高到95%。此外，向前和向左跌倒的识别准确率也提高了近14%。因此，实验结果表明，在不同场景下，四个跌倒方向的识别准确率提高后均超过95%。
