融合注意力机制的OpenPose人体跌倒检测算法  

OpenPose Human Fall Detection Algorithm Based on Attention Mechanism

作　　者：孟彩霞 薛洪秋[3] 石磊 高宇飞[3] 卫琳[3] Meng Caixia;Xue Hongqiu;Shi Lei;Gao Yufei;Wei Lin(School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001;Department of Image and Network Investigation Technology,Railway Police College,Zhengzhou 450053;School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450002)

机构地区：[1]郑州大学计算机与人工智能学院,郑州450001 [2]铁道警察学院图像与网络侦查系,郑州450053 [3]郑州大学网络空间安全学院,郑州450002

出　　处：《计算机辅助设计与图形学学报》2024年第12期2040-2050,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2020YFB1712401-1);国家自然科学基金青年科学基金(62006210);2022年度河南省重大科技专项(221100210100,221100211200);河南省研究生联合培养基地项目(YJS2023JD04);河南省重点研发与推广专项(科技攻关)(232102210095);南阳市协同创新重大专项(22XTCX12001);河南省高校青年骨干教师培养计划(2020GGJS256);河南省高校重点科研项目(24B520032);郑州警察学院基科费项目(2023TJJBKY010,2020TJJBKY006).

摘　　要：人员密集场所跌倒事件易引发公共安全问题,对人体跌倒进行实时监测和预警可降低安全风险.针对现有基于姿态估计跌倒检测方法模型规模大、时效性差等问题,提出一种融合注意力机制的OpenPose人体跌倒检测算法DSC-OpenPose.首先借鉴DenseNet稠密连接思想,将每层与之前所有层在通道维度上直接连接,实现特征复用,减小模型参数规模;然后在不同阶段之间添加坐标注意力机制,获取特征图空间方向依赖和精确位置信息,提高姿态估计精度;最后提出一种基于人体外椭圆参数、头部高度、下肢高度共同识别跌倒行为的方法,实现人体目标的跌倒检测.实验结果表明,在COCO数据集上,所提算法在模型规模和精度之间取得了较好的平衡效果;在real fall(RF)数据集上,所提跌倒检测算法的准确率达到98.2%,精度达到96.6%,检测速度达到20.2帧/s,且模型规模较小,满足嵌入式设备实时推理需求.Falls in crowded places are easy to cause safety problems,and real-time monitoring can reduce safety risks.Aiming at the problems of large scale and poor timeliness of existing fall detection methods based on pose estimation,this paper proposes an OpenPose human fall detection algorithm DSC OpenPose that integrates atten-tion mechanism.Drawing on DenseNet’s dense connection idea,each layer is directly connected with all previous layers in the channel dimension to achieve feature reuse and reduce the scale of model parameters.A method of identifying fall behavior based on the outer ellipse parameters,head height and lower limb height is proposed to realize the fall detection of human objects.The experimental results on COCO dataset show that compared with other algorithms,this algorithm achieves a good balance between model size and accuracy.At the same time,the fall detection method proposed in this paper has an accuracy of 98.2%,a precision of 96.6%,and a detection speed of 20.2frame/s on the real fall(RF)data set,and the model scale is small to meet the real-time reasoning needs of embedded devices.

关 键 词：人体姿态估计 密集连接 注意力机制 跌倒检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]