The Research Of Wearable Human Falling Recognition Method

At present,China has showed that a rapid growth trend of old aging and empty-nested elderly people in the background of global aging structure and the elderly living alone. With the growth of the age, the ability of the elderly begins to decline. And the elderly easily encounters a variety of dangers in their everyday life. The elderly's accidental injury brings a heavy burden to social medical resources. So more and more attention be taken to the elderly's day-to-day health care. In order to avoid serious consequences when the elderly falls down in the ground with no help for a long time, it is important that the system of human falling recognition can accurately differentiate between fall events and active daily life events and send alarm signals in the cases of not affect the elderly's normal life.The wearable system of human falling recognition is a part of the health care system. The falling recognition algorithm is the key part of this system. Since one misjudgment of falling activity may cause much more serious damage. Because the threshold method has only one feature and largely relies on experience point. Based on acceleration and human inclination, a falling recognition algorithm of SVM(support vector machine, SVM) is put forward after analysing the characteristics of human motion. The main research contents are as following:(1)The basic principle of SVM is particularly explained. And SVM classifier is designed by marking samples, normalizing samples, training SVM, revising the threshold of marking samples. In order to deal with the problem of unbalanced sample, this paper puts forward Biased-SVM. By comparing the recognition performance between inverse proportion and average density in the selection method of two different penalty factors, and the average density achieves relatively good results.(2)To further improve the recognition performance, over-sampling is put forward to deal with unbalanced samples from the data layer. In order to solve the problem of noise samples and overlapping samples by using SMOTE(synthetic minority over-sampling technique, SMOTE), a new algorithm of CDB-SMOTE(center distance-based synthetic minority over-sampling technique, CDB-SMOTE) is put forward to avoid blinding to generate samples. Based on the sparse degree of fall samples, the distribution function can be gotten by calculating the distance between subsample and the center point of the class. This algorithm gets new samples which are in favor of the falling recognition by using distribution function.(3)This paper designs and realizes a visual simulating software of the falling recognition algorithm with the software platform of VS 2010. To verify the feasibility of Biased-SVM algorithms based on CDB-SMOTE, the samples are tested by using Biased-SVM classification model. The test result proves that the optimized falling recognition algorithm of Biased-SVM which meets the need of system's reliability and accuracy is better than other algorithms.