改进TCN算法在人体跌倒检测中的应用  

Application of improved TCN algorithm in human fall detection

作　　者：魏嘉雪 高冠东[2] 滕桂法[1,3] WEI Jia-xue;GAO Guan-dong;TENG Gui-fa(College of Information Science and Technology,Hebei Agricultural University,Baoding 071000,China;Department of Information Management,The National Police University for Criminal Justice,Baoding 071000,China;Hebei Key Laboratory of Agricultural Big Data,Hebei Agricultural University,Baoding 071000,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071000 [2]中央司法警官学院信息管理系,河北保定071000 [3]河北农业大学河北省农业大数据重点实验室,河北保定071000

出　　处：《计算机工程与设计》2023年第9期2859-2866,共8页Computer Engineering and Design

基　　金：河北省自然科学基金项目(C2020204055);河北省教育厅科学研究重点基金项目(ZD2021056)。

摘　　要：为提高跌倒检测的准确率,解决传统RNN和CNN训练模型复杂且易产生梯度爆炸现象的问题,提出一种改进的时间卷积网络(TCN)算法。借鉴ResNet恒等映射的思想对残差结构进行改进,将激活函数改进为Leaky ReLU,减少神经元坏死的现象,为避免参数冗余造成模型过拟合问题,选用全局平均池化层代替全连接层实现分类。实验结果表明,该算法判断准确率达到99.4%,较改进前提高了10.51%,与其它已有算法相比准确率提高了2.68%~3.63%,能够准确检测出跌倒行为,对于及时识别老年人跌倒并报警,预防因发现不及时致残致死,具有较高的实用价值和社会价值。To improve the accuracy of fall detection and solve the problem that traditional RNN and CNN training models are complex and easy to produce gradient explosions,an improved time convolutional network(TCN)algorithm was proposed.The idea of ResNet identity mapping was used to improve the residual structure.The activation function was improved to Leaky ReLU,which reduced the phenomenon of neuron necrosis.To avoid the problem of model over-fitting caused by parameter redundancy,the global average pool was selected.The classification layer was used to replace the fully connected layer to achieve classification.Experimental results show that the accuracy of the algorithm’s judgment is 99.4%,which is 10.51%higher than that before the improvement.Compared with other existing algorithms,the accuracy is increased by 2.68%~3.63%.It can accurately detect the behavior of falling and is useful for timely recognition.For reducing the disability and death caused by untimely discovery of the elderly fall,it has high practical and social value.

关 键 词：时间卷积网络 跌倒检测 残差结构 恒等映射 激活函数 全局平均池化 参数冗余 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]