Research On Living Behavior Recognition System For Elderly Living Alone Based On Deep Learning

As our country enters the deep aging society,the life condition of the elderly has been paid more and more attention,especially the life condition of the elderly living alone whose family is not around and the activity ability is weak should be paid more attention.In order to enable the guardian to obtain the living behavior information of the elderly living alone at any time,it is necessary to study a living behavior recognition system for the elderly living alone.The system designed in this thesis adopts joint node behavior recognition technology for behavior recognition of elderly people living alone,and then uses Raspberry PI,server and Android phone to build a complete system.The main research contents include:(1)Spatiotemporal graph convolutional neural network(ST-GCN)was improved into a dual graph fusion network.The attention mechanism was improved by using mask addition instead of mask multiplication,and then the aspect ratio of human body was reduced to the aspect ratio of joint nodes in six regions,which was named joint domain.Then,joint domain features are constructed into joint domain graph and feature extraction is carried out.Joint node feature extraction branches and joint domain feature extraction branches are fused to form a double graph fusion network.On the self-made dataset,the accuracy of dual-graph fusion network is 98.9% and the number of model parameters is 5.2MB,while the accuracy of ST-GCN is 97% and the number of model parameters is 12.5MB.Therefore,the dual-graph fusion network is better.(2)The production of life behavior data sets.A total of 4198 short videos simulating the walking,falling,eating,sleeping,getting up,sitting and standing behaviors of elderly people living alone were filmed and made into a behavior recognition dataset.In order to improve the efficiency of obtaining short videos and making label documents,Tkinter library was used to write the labeling software and automate the labeling of the dataset.(3)Design of living behavior recognition algorithm for the elderly living alone.Then the motion detection algorithm was designed according to the coordinates of the joint nodes.Then the fall detection algorithm,meal time statistics algorithm and sleep time statistics algorithm were designed by combining BlazePose,motion detector and dual graph fusion network.These algorithms make the system more applicable to practical scenarios.(4)Design and implementation of living behavior recognition system for the elderly living alone.Raspberry PI was used to build the indoor terminal,the server composed of Flask framework and My SQL database,and the Android phone was used to build the living behavior recognition system for the elderly living alone for the mobile terminal.Raspberry PI is responsible for collecting indoor real-time image information and is responsible for the functions of human pose estimation,motion detection,email alarm and communication with the server,while the server realizes the functions of human behavior recognition,data persistence operation and communication with Raspberry PI and Android APP.The Android APP is responsible for obtaining the behavior information of the elderly living alone from the server and visualizing it.