摘要
针对独居老人摔倒问题,构建一种基于双流卷积神经网络(TwoStream CNN)的实时跌倒检测模型。将提取人物轮廓的RGB单帧作为输入的空间流,将连续多帧运动历史图(motion history image,MHI)作为输入的时间流融合,在一个特定维度的全连接层将两个网络的同shape张量Concatenation相连后添加到新的层,在公开的摔倒数据集上进行实验和定量分析。实验结果表明,采用人物轮廓RGB-MHI的双流卷积网络在非摔倒和摔倒的区分检测中准确率达到了98.12%,改进的融合方式相比较其它方法有提高,时间流输入MHI满足实时性要求。
Aiming at the fall problem of elderly people living alone,a real-time fall detection model based on two-stream convolutional neural network(TwoStream CNN)was constructed.The RGB single frame of the person’s outline was taken as the input spatial stream and the continuous multi-frame motion history image(MHI)was used as the input time stream.The same shape tensor Concatenations of two networks were connected using the fully connected layer of a specific dimension,and then added to a new layer.Experiments and quantitative analysis were performed on the public fall data set.Experimental results show that the TwoStream CNN using the character contour RGB-MHI has an accuracy rate of 98.12%in the non-fall and fall detection.The proposed fusion method is improved compared to other methods,and the time stream input MHI meets real-time requirements.
作者
金彦亮
陈刚
JIN Yan-liang;CHEN Gang(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)
出处
《计算机工程与设计》
北大核心
2021年第9期2621-2626,共6页
Computer Engineering and Design
基金
上海市科委重点基金项目(19511102803)。
关键词
跌倒检测
双流卷积
卷积神经网络
运动历史图
模型融合
fall detection
two stream convolutional
convolutional neural network
motion history map
model fusion
