A Fall Detection System Based On Android Phones

The fall is one of the important factors of hazarding the elderly and other special populations. Promptly fall detection and rescue can gain precious time for treatment and salvage, and it is very important to secure health of people and improve quality of medical care. At present, our country has entered the aging society, which presents accelerating development tendency as well. So portable, accurate and real-time fall detection method and system urgently need to be researched and developed, in order to meet wide social requirements.Currently, the domestic and international numerous research institutes and universities have studied and experiment fall detection method continuously, but existing methods generally train One-Class Classification model by feature extraction from the normal activity data, such as walk and running. However, the transient feature information of individual normal activity such as running and down stairs is highly similar to fall, and with the influence of the noise, which leads to distinguish ability of the model is insufficient and can’t meet the requirements of high detection accuracy and low false alarm rate.In order to improve the practicability of fall detection method, this paper, by statistical analyzing the characteristics of fall occurrence, proposes a fall detection method based on activity transition. The research of this paper mainly includes the following four parts:1、Highly accurate continuous activity recognition model. It builds a continuous activity recognition model to identify the human activity. This stage comparatively analyzes the performance of diverse multiple classifier by a large number of experiments, and verifies that the multimodal sensors detection result performs better than single sensor. It filters feature set in accordance with the feature selection function of sample anlysis tool, and obtains more robust feature set;2、Automatic segmentation technology for activity transition data. It segments continuous activity sequence according to the result of activity recognition, and obtains activity transition dataset. This paper proposes an automatic segmentation method for activity transition data;3、Research and construction of anomaly detection model. It builds anomaly detection model by feature extraction from the normal transition dataset of adjacent activities. The model detects fall by identifying abnormal activity transition. The activity transition based fall detection method builds feature space by activity transition dataset, which can filtrate numerous normal activity date, reduce complexity of feature space, and enhance the distinguish ability of model. Comparative experiments of this paper show that classification accuracy of our method is better than the traditional methods, which has certain practical value.4、A fall detection system based on Android phones is designed and implemented. The system can collect real-time sensor data, and real-time monitor the people falls or not by the fall detection model. If the people fall, mobile phone will alarm with buzzer sound locally, and obtain the location information of the people by localization module, and automatically edit the fall information and location information to create SMS messages, and then sent it to the specified contact in order to get timely rescue.