Research On Fall Detection System Based On Deep Learning

The proportion of the ageing population in China is increasing,and falls are one of the main causes of injuries among the elderly,and it is important to provide assistance in the first instance for the safety of the elderly.In order to improve the accuracy of fall detection and provide medical assistance to the elderly who have fallen at the first time,this paper designs a fall detection system based on deep learning.The main research content and innovative work are as follows.(1)In this paper,a wearable fall sensing system is designed to measure three-axis acceleration and three-axis rotational angular velocity.Use the bluetooth module to transmit the data to the upper computer.In addition,considering the wearing comfort and stability,a human motion model was established,and the right waist was selected as the characteristic part after comprehensive analysis.Finally,the various modules of the hardware are jointly debugged to realize the flow of information.When the system finds a fall,it will automatically send an alarm text message to the designated number through the cloud platform alarm system.(2)To address the problem of low detection accuracy caused by the long sequence characteristics of fall detection data information,this paper first pre-processes the data,processing the long sequence data into spatio-temporal data,thus achieving the purpose of reducing the length of data in a certain dimension and preventing the network from going out of order.A time window is also used to obtain the data set in order to prevent the moment when the fall occurs from not being in the data set.(3)In order to solve the problem that fall detection is easy to distinguish and misjudge,this paper proposes a fall detection algorithm based on CNN-Casual LSTM network for fall detection judgement to improve the accuracy of fall detection.The neural network mainly consists of an encoding layer,a decoding layer and a Res Net18 classifier.The coding layer consists of three layers of CNN and three layers of Casual LSTM.the decoding layer consists of three layers of deconvolution and three layers of Casual LSTM.the decoding layer maps spatio-temporal information to hidden variable outputs that are more conducive to the work of the classification network,which is classified by Res Net18.This spatio-temporal network model makes full use of spatio-temporal information by extracting the spatial features of each input through the coding layer,while retaining the influence of temporal information through the GHU fast channel,with a higher recognition rate of fall behaviour.(4)Aiming at the number of layers of the network structure and the selection of neurons,this paper verifies the rationality of the network structure setting by means of the control variable method.In addition,algorithm performance is evaluated in multiple dimensions such as accuracy by using public datasets and algorithms such as FD-CNN for comparative experiments.The experimental results show that the ACC of the public dataset is 99.79%,SEN is 100% and SPE is 99.73%,while the ACC of the actual measured dataset is 98.33%,SEN is 100% and SPE is 97.50%.