Research Of Multi-feature Fusion Method Based On Fall Detection

As the global aging process quickened, the human fall detection, warning and protection technology has become an important guarantee for the elderly, preventing the accidental falls injury and improving the quality of people’s lives. Meanwhile, because of the development of wearable computing and body sensor network technology, fall detection has become a new research. This technology not only has high significance in the study of theoretical, but also in practical applications.Currently, the fall detection technology in domestic and international research is only limited to identify the accident after a fall which can help reduce secondary damage by whistle or sending information to someone. Fall warning is different from fall detection which can provide an early warning before the fall and take appropriate methods, such as airbags to protect the body from hurt as far as possible.The research of fall detection is divided into four parts, building the model of falls, designing experimental, Processing of the raw data and applications of algorithms. Acceleration and angular velocity are the main features, with threshold method, support vector machines, decision trees and other machine learning algorithms to realize fall detection or warning. Fall detection is essentially a simple binary classification problem, what we need to do is to distinguish ADL from fall. Threshold is the most common research methods and easy to realize. At present, there is no maturity model. The threshold value is fixed mostly based on statistical experience and experimental results. It has a great impact on the recognition results by different threshold values, which considered poor stability. Pattern recognition method has a great advantage for large sample classification, but in the actual research, it is difficult to obtain a large number of real falls data Support vector machines has some advantages for solving nonlinear small sample and pattern recognition problems.This paper is committed to a falling detection through self-adaption multi-feature fusion. First, analyse the output data of nine axis inertial sensor, select two features just as acceleration and angular velocity that distinguish fall activities from ADL; constructed the virtual samples of falls and daily activities using gaussian distribution principle; Then, set up single SVM model for weight vector, realize multi-feature fusion of the acceleration and angular velocity through feature weighted kernel function; Finally, realizing fall detection by the use of multi-feature fusion algorithm based on SVM.The results show that the best location to worn sensor is the body’s side of waist. In the fall detection model based on SVM, polynomial kernel has an excellent performance, while multi-feature fusion algorithm based on SVM can distinguish fall activities from daily behavior effectively.