基于双向循环神经网络的跌倒行为识别  

Fall behavior recognition based on bidirectional recurrent neural networks

作　　者：佃松宜[1] 程鹏 王凯[1] 雒瑞森 DIAN Song-yi;CHENG Peng;WANG Kai;LUO Rui-sen(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电气工程学院,四川成都610065

出　　处：《计算机工程与设计》2020年第7期2019-2024,共6页Computer Engineering and Design

摘　　要：针对基于三维人体特征点识别跌倒行为需要专用相机设备的问题,提出一种基于二维人体特征点的跌倒行为识别方法。不需专用的相机设备支持,使用开源的计算机视觉库从RGB视频帧提取二维特征点,双向循环神经网络对特征序列进行识别,使用门控循环单元作为循环神经网络的循环单元,变分丢弃法作为网络的正则化项。实验结果表明,与新的跌倒专用数据集CMDFALL的基准算法相比,该方法在节省成本的同时提高了精度。To solve the question that fall behavior recognition based on three-dimensional characteristic points of human body needs focused camera devices,a method based on two-dimensional characteristic points of human body was proposed for fall behavior recognition,which did not need the support of focused camera devices.An open-sourced computer vision library was used to extract two-dimensional characteristic points from input RGB video frames.Characteristic sequence was inputted to bidirectional recurrent neural networks for recognition,in which gated recurrent unit was chosen as the recurrent unit and variational dropout was used for regularization.Experimental results show that the proposed method not only saves cost,but improves the accuracy compared with benchmark algorithm in the new fall focused dataset CMDFALL.

关 键 词：跌倒行为识别 二维人体特征点 双向循环神经网络 门控循环单元 变分丢弃法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]