Human Fall Recognition Method Based On Mechanical Information Acquisition System

Fall accident of elders is always a serious problem in social community. As the world aging process quickened, it has became a significant financial burden to modern society. Hence, reliable fall prediction and detection method is essential to reduce fall related injury in independent living facilities, and it has become a hot investigation world wide. However, there is not special model for human fall process research nowadays, so as to lead the misdetection of falls and similar motions, especially for fall prediction methods. Thus, there are still many key issues unsolved.This paper is committed to explore human fall detection and prediction method through the theories of mechanical information acquisition, so as to improve the detection accuracy and prediction real-time property. Its main works and contributions are summarized as follows:(1) Through analyzing the sports biomechanics features of human fall process, human upper trunk is defined as the feature region to extract information for fall recognition. And accordingly, variable acceleration and the tilt angle deviate from the vertical orientation of human upper trunk are chosen as the features for fall detection and prediction study.(2) Classify the resultant accelerations and tilt angles deviate from the vertical orientation of human upper trunk from fall processes and other daily life activities using Support Vector Machine (SVM), and take the acceleration and angle value on the optimal classification boundary as the thresholds to detect fall evens. Through analyzing the relationship between acceleration threshold, angle threshold and time, two human fall detection methods are proposed and realized: one method is based on tri-axial accelerometer and bi-axial gyroscope, the other one is based on tri-axial accelerometer.(3) To detect and predict fall evens, a human motion states timing analysis based human fall recognition method is proposed. First, extract the acceleration time series (ATS) through human motion process to characterize the process. Then, take human motion movement as random process, to study the transition probabilities between each motion state in human fall process, and the appearance probabilities between motion states and ATS. Because fall prediction result must be decided before the collision during falls, this paper build the Hidden Markov Model (HMM) to describe human motion particularly during falling courses but before the collision between human body and lower objects. Hence, the output probability of ATS, which is from a period time before current time during a motion process, on the HMM represents the marching degree between that motion process and HMM, it can evaluate the risk to fall down at current time, so as to realize the fall detection and prediction method. At last, the author got the output probability thresholds between fall process and other daily life activities through SVM, and designed human fall detection and prediction algorithm.(4) To practice experiments of human fall detection and prediction algorithms, an experimental platform with information acquisition system for human fall recognition is build. In the limit of experimental samples, the human motion states timing analysis based human fall recognition method in this paper can distinguish fall evens and other daily life activities effectively: 100% sensitivity and 0% specificity (no misdetection) is got in detection experiment, and the prediction time interval is 200~300ms before the collision of human body with lower objects. The results of this study above enriched the knowledge of human fall detection and prediction study based on the previous research.