The Design Of Human Motion Detection System And The Study Of Fall Prediction Method

Fall accident is recurring often in elderly people’s daily life, and it brings serious physical and psychological harm to them. The differences between falls and other daily activities are the accelerated speed of human body motion and the angle change that deviates from the vertical direction. Domestic and international research on fall recognition and prediction are diversified. Considering fall recognition effectiveness and protecting personal privacy, this paper introduces a design of a wearable, fall detection system which extracts the3axis accelerated speed of human chest, as well as the angular speed of the pitch angle and the roll angle, as the features of fall accidents.The main work of the paper includes:1) The hardware and software modeling of fall detection system. An architecture model of the information acquisition terminal, which is the portable part of the system, is built based on AADL in OSATE environment. In the meantime, the end to end flow latency analysis tool in the OSATE is used to estimate the real-time response performance of the terminal so as to determine the appropriate sensor sampling frequency.2) The design and implementation of the information acquisition terminal. The terminal is able to acquire human kinematics parameters, detect fall accidents and send alarms to the remote base station. In the perspective of hardware design, the terminal comprises a microprocessor, an accelerometer, a gyroscope, and a wireless transmission module; while the software subsystem is responsible for acquiring and processing real-time sensor data, then using the detection algorithm to determine whether the fall occurs, and finally sending the alarm to the remote base station if the algorithm recognizes the fall.3) An optimized fall detection algorithm is proposed based on support vector machine. The resultant acceleration velocity and the resultant angular velocity are acquired as features when the sampling data shows a peak and its next sampling point, and trained to build a fall detection classifier. The classifier is able to distinguish the human fall from other daily activities. The effectiveness of the classifier is cross-validated via the LeaveOneOut method. After that, the algorithm is optimized by adding threshold analysis in order to reduce misjudgment. Experimental data shows that the sensitivity and specificity of the optimization is better.4) Fall prediction algorithm is realized based on Hidden Markov Model. The features before fall occurring and other daily activities are extracted and integrated into angular velocity time series. Then these two kinds of time series are trained to build the prediction model and other daily behavior model, respectively. Finally, the category of each observation series is determined via its matching probability with these two models. Therefore, the human fall can be predicted and the prediction accuracy is between0.2s and0.6s via the experimental result.