Human Fall Detection Method Based On MEMS Sensor

Since China entered the aging society at the beginning of this century,the aging population of our country has been deepened,the number of elderly people has been increasing,and the number of people over 60 or 65 years old is huge.How to deal with the problem of population aging and protect the life and health of the aging population has become a focus topic in society.Accidental falls of the elderly occur frequently,which is likely to cause adverse physical or psychological consequences.It is also one of the main causes of illness,disability and death of the elderly,which seriously affects the life and health of the elderly.In order to reduce and avoid the negative impact of falls on the elderly,in addition to targeted fall prevention,family members or guardians are required to detect and treat the elderly's fall behavior in a timely manner.Therefore,it is of great significance to study the method of human fall behavior detection to reduce the negative impact of accidental falls.This thesis focuses on human fall behavior detection methods,and mainly performs the following work:Analyze the advantages and disadvantages of the existing fall detection methods,and determine the nine-axis MEMS sensor as the basis for data collection.By analyzing the biomechanical characteristics of the human body during the fall process,a MEMS sensor was worn on the waist of the upper torso of the human body,and a data set containing falls and six other non-falling human behavior data was collected.For the data set used in this paper,a data preprocessing method is given.Extract the original data from the data set,store it in an easy-to-use form,use Kalman filtering to remove abnormal data from the MEMS sensor data,and smooth the data,which is conducive to more accurate subsequent experimental results.Aiming at the problem that the threshold detection method is poorly adaptable to complex and diverse user fall behavior characteristics,a human fall detection method based on an improved adaptive multi-threshold model is proposed to improve the adaptability of human fall detection.The main content of the research is to set up a system based on the target person's own multiple behavior feature extraction and based on the human body's fall time multiple state conversion system.In addition,volunteers with different BMI values were invited to simulate complex and diverse users for experimental verification,and the fall detection accuracy rate reached 99.22%.Aiming at the problem that the recurrent neural network has the gradient disappearance and it is difficult to learn long-term dependencies when using deep neural network to recognize the fall behavior,a human fall detection method based on improved LSTM of Kalman filtering is proposed.Ability,learning ability of long-term memory,short-term memory,unique gate structure,and Kalman filter's ability to estimate the state of dynamic data,process and analyze sensor data to identify 7 normal behaviors and 9 different fall behaviors in the experimental comparison,the recognition ability of the method was verified,the average accuracy rate reached 98.84%,and the accuracy rate of human fall behavior detection was 98.73%.