Research On Fall Detection For Elderly Based On Classification

With the development of social aging,falling down has become a common health problem among the elderly,and its detection has become an essential direction to study the health problems of the elderly.There are high-risk falls for hospitalized elderly patients in toilets and hospital beds based on the analysis of clinical elderly tumbling data.However,these particular scenes with different needs characteristics become a research problem involving the elderly's privacy.In this paper,Kinect body sensor is used to obtain and process data such as human depth image and audio signal and perform fall behavior detection through multi-source heterogeneous data combination.This thesis studies the falling detection of the elderly in specific indoor scene from toilet,night-time geriatric ward,and day-time geriatric ward.A classification learning method is introduced to carry out the falling detection of the above scenes.The work is carried out as follows:1.In the detection of the elderly's falling behavior in the toilet,where is a high-risk place for the elderly to fall accidentally.In this scene,the falling behavior occurs frequently due to slippery floors or fainting in the elderly,and family members or caregivers cannot detect it in time.Due to that the wearable sensor is not waterproof and the video sensor invades privacy,this paper proposes to use Kinect sensor to obtain audio signals in this scene.And perform blind source separation on the audio mixed signal by independent component analysis models.Finally,a falling behavior detection will be performed by threshold method based on energy characteristics in short time,and the experimental results obtained an accuracy rate of 85.00%on average.2.The lack of lighting in the night-time geriatric ward makes the elderly extremely prone to falls.Traditional vision-based methods rely on tracking skeletons to estimate height changes in key parts of the body(such as the head,hips,etc.).It is often challenged by posture changes and of body parts shielding.The paper proposes(a Patical Swarm Optimization Extreme Learning Machine-Convolutional Neural NetWorks,PsoELM-CNN)to detect human fall behavior based on deep images of human daily behavior.The experimental data includes 32 falling behavior and 8 daily behaviors,with a recognition rate of 99.82%.3.In the detection of the fall behavior of the elderly in the daytime geriatric ward,it is bright at daytime in geriatric ward with limited activity range,the thesis uses wearable bracelet and Kinect to obtain human triaxial acceleration data,depth image and skeleton coordinate data out of consideration of personal privacy and frequent activities of the elderly.After preprocessing and feature extraction for the three types of heterogeneous,decision-level fusion is used to identify the fall behavior finally.Skeleton node data is preprocessed by frame sampling,node processing,and coordinate normalization at first and CNN-LSTM model is used for feature extraction.Depth image data is converted into motion history image during preprocessing for being normalized and feature extraction via CNN-LSTM model.Tri-axial motion coordinates are converted into acceleration data and classified by SVM.Finally,the decision-level fusion is used to determine the fall behavior of the three heterogeneous data through a softmax.The accuracy rate is 95.80%in data model training.4.The human behavior data acquisition system is obtained based on WPF framework and Kinect for Windows SDK v2.0 interface,with multi-functional design such as data type selection,frequency,duration,data storage path,etc.The mobile terminal monitoring platform will be realized based on the WEB framework,Eclipse tools and Navicat database and the elderly's fall monitoring and rescue fuction will be achieved.