Research On Fall Detection Based On Wearable Sensor

Falls are considered to be one of the most serious health problems leading to the burden of disease in older people.Strengthening the elderlys fall intervention is conducive to the construction of a health information service system.The focus of research on fall intervention is fall detection.It is considered to have the advantages of cost and practicality in fall detection by sensors.This thesis analyzes the core issues of wearable fall detection.It provides a technical basis for the fall detection by the sensor.Fall detection is discussed in this thesis.A single sensor is used to collect activity data and detect human falls.Multiple wearable sensor units are used to detect falls.Analyzing the combinations of multiple wearable sensing units.Analyzing the difference between a real fall scene and a simulated fall scene.It detects fall in certain scenes.The thesis work is as follows:1.By combining sensor units such as accelerometer and spiral meter,the human body can be obtained in multiple dimensions.The fall feature is extracted from the collected data sequence,and then the dimension of the feature is reduced.Then,the incremental learning algorithm Learn++ classifies the features after dimensionality reduction.Learn++ has sub-classifiers,using support vector machines and C4.5 decision algorithms.The sub-classifier generated based on the users personal data is combined with the public sub-classifier to generate a personalized fall detection algorithm.In the SisFall data set,the accuracy of the Learn++ algorithm is 91.11%.2.It performs fall detection by wearing sensing units at multiple positions on the human body.The sensors are placed on the chest,waist,wrist,ankle,etc.Convolutional Neural Networks are used to extract the features of the data.Long Short-Term Memory models are used for classification.The results show that the detection accuracy of multisite combination classifier is higher than that of single-site.The combined classifier detection rate for the chest,waist,wrist,and ankle was the highest,which was 88.57%.The detection accuracy through the chest data is 80.00%.3.There is a difference between the actual fall scenario and the simulated fall scenario,which leads to a decrease in the detection rate of the fall detection algorithm during application.Simulated complex falls more similar to real falls were discovered.And a near fall that is easy to be misjudged was discovered.Through the Long Short-Term and Memory-Full Convolutional Neural Network,these fall features can be automatically extracted.When detecting complex falls and near falls,the accuracy of the algorithm reaches 98.71%.It was found that through more complex falls and near falls,it is closer to the real fall environment.4.The fall detection and analysis platform is designed and implemented.The platform is divided into mobile and platform.The function of the mobile terminal is fall detection,which displays real-time information of personal signs to the user.The platform-side function is the integrated effect of the analysts wearable sensing unit,which displays and counts historical fall record information.