Early Warning System for Physical Distancing Detection in the Prevention of COVID-19 Spread
Early Warning System for Physical Distancing Detection in the Prevention of COVID-19 Spread
复制标题
用于预防 COVID-19 传播的物理距离检测的早期预警系统
DOI：
10.1109/icodsa53588.2021.9617553
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
V. Suryani
中科院分区：
文献类型：
--
作者：
Abdullah Hadi;Rizka Reza Pahlevi;V. Suryani
COVID-19病毒的广泛传播对人类健康构成威胁;因此，预防COVID-19传播需要立即实现。物理距离是减少COVID-19传播的一种方法。然而，由于缺乏严格的监督，在实施物理距离方面经常发生人为疏忽。许多研究，其中之一是基于相机机器学习的系统，试图解决这个问题，但他们专注于检测精度，而没有考虑实际上需要检测物理距离的设备移动性。我们提出了一个系统，可以提供早期警告，对物理距离疏忽的限制计算机。该系统使用受限的计算机和具有高移动性的相机来构建，以便于其移动。该系统使用Tensorflow Lite作为机器学习的框架，并使用SSD MobileNet预训练模型对人体检测进行分类。应用的测试场景包括距离准确度和检测准确度。该系统在检测物理距离疏忽方面具有86%的准确率和87%的F-1得分。该系统可以在有限的计算机上运行，使用4.59 MB的内存，占总内存的0.001%，四核的累积利用率为139%。该系统的精度比同类工作提高了10%。
The widespread of the COVID-19 virus poses a threat to human health; hence, the prevention of the Covid-19 spread needs to immediately be realized. Physical distancing is a way that can reduce the COVID-19 spread. However, human negligence in implementing physical distancing due to the lack of strict supervision often occurs. Many studies, one of which is Camera Machine learning-based system, have attempted to solve this problem, but they focused on detection accuracy without considering device mobility that, in fact, is needed to detect physical distancing. We proposed a system that can provide early warnings against physical distancing negligence on the constrained computers. The system was built using a constrained computer and a camera with high mobility to facilitate its movement. The system used Tensorflow Lite as a framework to do machine learning and the SSD MobileNet pretrain model was used to classify human detection. Test scenarios applied included distance accuracy and detection accuracy. The system had accuracy in detecting physical distancing negligence with 86% accuracy and 87% F -1 Score. The system built can run on a constrained computer by used 4.59 MB of memory that is 0.001 % of total memory, and the cumulative utilization of four cores was 139%. The system performs 10%better in accuracy than a similar related work.