| Fall detection systems aims at classifying falls from activities of daily life, and warns people of danger in order to shorten the rescue time and reduce the hazard of falls. It should be provided with social significance that to carry out study of fall detection in the background of aging population and old people living separately with their adult children. This paper designs algorithm and experiments according to four parts of body(chest, side waist, back waist and thighs) which are frequently used to wear the mobile phone. This article has developed a cell phone application for fall detection and a server, which can use to falls recognition and give an alarm. The study consists of three parts:The first is the experimental design. First, we carry out pattern classification of human activity. Then, we design the experiment, including experimental pattern of activity, experimental site layout, the choice of subjects and specific experimental procedures. Finally, four inertial sensors were worn four positions of wearing the mobile phone to record the kinematic and posture data of experimenter. We get the raw data to design and test algorithm by this experiment.The second is the algorithm design. We design feature extraction method and back-propagation neural networks by in-depth analysis of the experimental data and process of falls. We consider six user habits of wearing phone(four kinds only wear a fixed body parts(chest, side waist, back waist or thighs), one kind wear in three parts of the upper-body(chest, side waist, back waist), one kind wear four parts of body(chest, side waist, back waist and thighs)) and two kind mobile phone(whether with gyroscope or without gyroscope) to design twelve kinds of neural network model. We draw the following conclusions:(1) It will appear many false negative and false positives when a phone without gyroscope wear in thigh.(2) The sensitivity and the specificity is 100% and 100% respectively when wear the phone with gyroscope in three parts of the upper-body(chest, side waist and back waist). The sensitivity and the specificity is 98.67% and 96.54% when wear the phone without gyroscope in three parts of the upper-body. The sensitivity and the specificity is 96.12% and 97.17% when wear the phone with gyroscope in four parts of the body(chest, side waist, back waist and thighs). The sensitivity and the specificity is 89.28% and 95.07% when wear the phone without gyroscope in four parts of the body. The sensitivity and specificity of the algorithm of this paper is the highest by compare with several similar researches.(3) We adopt two kinds of neural networks models for falls detection, which is a neural networks model of four parts of the body and a neural networks models of three parts of the upper-body respectively. Users can switch between the two kinds of neural networks modelsThe third is the system design and realization. The system includes an android-based application for falls detection and a back-end server based MFC and SQL SERVER database. The falls detection algorithm develops by OpenCV which call by JNI. Cell location based on Baidu Maps SDK adopts one of the three positioning technology: GPS satellite positioning, base station positioning and Wi-Fi positioning when fall occur. Then, the system send an SMS with address of falls to preset contacts and ring alarm bells for help. The relevant information of falls and false positives will be uploaded to the backend server for algorithm research and statistics when the mobile application detects a fall. At the same time, mobile application will add corresponding data to training set, re-training the neural network for improved algorithm. Remote server is responsible for information storage and user management by SQL SERVER. Mobile application communicates with remote server by TCP/IP protocol and socket.In summary, this paper designs experiments according to four parts of body which are frequently used to wear the mobile phone. Then, we design twelve kinds of neural network models from the perspective of the user and mobile phone, which is corresponding to six users’ habits of wearing phone and whether phone with gyroscope or phone without gyroscope. According to the experiment, we adopt two kinds of neural networks models to fall detection, which is a neural networks model of four parts of the body and a neural networks models of three parts of the upper-body respectively. Finally, we develop a fall detection system using android-based mobile phone. This is a hopeful technology to monitor for falls in the background of aging population and old people living separately with their adult children. |
