基于多流卷积神经网络的跌倒检测算法  

Fall Detection Algorithm Based on Multi-Stream Convolutional Neural Network

作　　者：张陶 邬春学[1] ZHANG Tao;WU Chun-xue(School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《软件导刊》2023年第1期145-151,共7页Software Guide

摘　　要：为了更好地检测家居场景下老年人的跌倒问题,提出一种基于多流卷积神经网络的跌倒事件检测模型。该模型在传统双流网络的基础上,添加一个新的融合流网络组成三支流网络架构。融合流以时间流和空间流所提取的时空特征为输入,经由多模态融合模块将对应层的时空特征融合,并将得到的所有时空融合特征再经融合模块融合得到的多级时空融合特征。最后将三支流的输出进行加权平均分数融合得到最终检测结果。实验结果表明,相比传统双流网络,该方法有更高的准确率、召回率和F1值,平均达到95.8%、91.3%和93.5%,证明了该方法的可行性。In order to better detect the fall problem of the elderly in the home scene,a fall event detection model based on multi-stream convolutional neural network is proposed.Based on the traditional two-stream network,this model adds a new fusion-stream network to form a three-stream network architecture.The fusion flow takes the spatiotemporal features extracted from the temporal flow and the spatial flow as input,fuses the spatiotemporal features of the corresponding layers through the multimodal fusion module,and fuses all the obtained spatiotemporal fusion features through the fusion module to obtain multi-level spatiotemporal fusion features..Finally,the outputs of the three tributaries are fused with weighted average scores to obtain the final detection result.The experimental results show that compared with the traditional two-stream network,the method has higher precision,recall and F1 value,reaching 95.8%,91.3% and 93.5% respectively,which proves the feasibility of the method.

关 键 词：跌倒检测 多流卷积神经网络 特征融合 深度学习 

分 类 号：TP391.04[自动化与计算机技术—计算机应用技术]