面向医疗护理的视觉监控医院患者跌倒检测-《中国医学物理学杂志》

文章信息/Info

Title:: Medical care oriented visual surveillance of patient falls in the hospital

作者:: 孙颖1; 张吟龙2; 王鑫3; 曾子铭4; 1.中国医科大学附属第一医院重症医学科，辽宁沈阳 110001； 2.中国科学院沈阳自动化研究所，辽宁沈阳 110016； 3.沈阳建筑大学信息与控制工程学院，辽宁沈阳 110168； 4.深圳职业技术学院汽车与交通学院，广东深圳 518055

Author(s):: SUN Ying1; ZHANG Yinlong2; WANG Xin3; ZENG Ziming4; 1. Department of Critical Care Medicine, the First Hospital of China Medical University, Shenyang 110001, China 2. Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China 3. School of Information and Control Engineering, Shenyang Jianzhu University, Shenyang 110168, China 4. School of Automotive and Transportation Engineering, Shenzhen Polytechnic, Shenzhen 518055, China

Keywords:: fall detection video surveillance medical care deep convolutional neural network abnormal behavior analysis

摘要:: 提出一种面向医院患者的视觉监控跌倒检测算法，解决患者由于意外跌倒不能被及时发现的问题，为医护人员快速处理患者跌倒等异常行为提供必要的技术保障。方法：首先，基于深度神经网络模型检测监控图像中人体关节点（如肩部、肘部、腕部、胯部、膝关节等）在图像中的位置，再根据亲和度向量场模型提取人体骨架，最后计算患者躯干、腿部与地面的夹角作为判别性特征，判断监控区域内是否有患者出现意外跌倒。结果：实验结果表明，本文所提算法在实际的医院监护环境中的处理速度高达25帧/s，检测准确率高达96%。结论：该方法能够实时、准确地提取医院环境下患者的行为特征，并针对意外跌倒情况发出警报，为医护人员监测患者跌倒等异常行为提供更准确、方便的计算机辅助医疗护理方法。

Abstract:: Objective To propose a visual surveillance algorithm for the detection of patient falls in the hospital, thereby solving the problem that hospital patients cannot be rescued in time when they fall accidentally, and providing the essential technical support for the medical staff to quickly deal with the abnormal conditions such as patient falls. Methods The positions of human joints (such as shoulder, elbow, wrist, hip, knee, etc.) in the image were detected based on deep convolutional neural network model, and the human skeleton was extracted using part affinity fields. Finally, the angles between the trunk or the leg and the ground were calculated as features to distinguish whether there were patients in the monitoring area who fall accidentally. Results The experimental results show that the processing speed of the proposed algorithm in the actual hospital surveillance environment was as high as 25 frames per second, and that the detection accuracy was up to 96%. Conclusion The proposed method can accurately extract the behavior characteristics of hospital patients in real time, and issue an alarm for accidental falls, providing a more accurate and convenient computer-assisted medical care method for medical staff to monitor abnormal behaviors such as patient falls.