基於行為狀態分析之跌倒檢測與預測之研究

A Study of Fall Detection and Prediction Based on Behavior Analysis

李鎧任(Kai-Ren Lee)

指導教授 ： 江季翰

國立虎尾科技大學/電機資訊學院/資訊工程研究所/碩士(2014年)

摘要

本論文提出一個能分析出人體日常運動行為的系統，且可識別出跌倒事件的發生。結合智慧型手機與手機內加速度感測器研究並實作系統，基於隱馬可夫模型(Hidden Markov model, HMM)的架構下，建置出行為狀態模型，透過該模型統計機率的特性，達到檢測各項行為狀態與識別跌倒的發生，研究方向主要以採樣頻率、樣本數量進行探討，最後為了提高系統識別的準確率，將提出了雜訊問題進行分析研究。
系統主要由兩個模組組成，訓練模組與檢測模組。訓練模組主要為了建立出行為狀態模型，首先我們觀察人體的運動行為，將設備放置於人體胸前，利用加速度感測器採集加速度值並建立出各項行為狀態樣本，之後藉由各項行為樣本分析出適合描述行為的特徵值，再提取特徵向量進行模型的參數訓練，並透過轉移矩陣描繪出狀態之間的轉移，最後將所有狀態串聯起來，建立出行為狀態模型；檢測模組主要進行行為狀態的識別，在檢測過程中利用維特比演算法計算狀態可能發生的機率，以推斷出實際發生的運動行為，達到該研究檢測目的。
實驗中將進行採樣頻率分析與訓練樣本數量比較，採樣頻率將比較在不同採樣頻率下所影響各種狀態識別的準確率，進而找出適合各項狀態檢測的頻率大小，而實驗結果在跌倒檢測上擁有93.9%敏感度與93.7%明確性，也說明該模型的有效性。在訓練樣本數量上，分析樣本數量所影響模型參數收斂度。實驗結果表示，此系統擁有識別不同的行為狀態，並提出各項狀態建議頻率大小與樣本數量。未來能藉由此系統檢測事件，在檢測出銀髮族發生跌倒後，能與智慧型手機通訊服務結合，即時地通報相關人員，降低跌倒發生的風險。

關鍵字

跌倒偵測 ； 跌倒預測 ； 隱馬可夫模型 ； 居家照護 ； 三軸加速器

並列摘要

Because of the gradual degradation of health, lots of risks occur in daily living of aging people. Among of them, the injuries caused by falls are the most common serious accident. In recent years, the topic about fall detection has been widely discussed in the area of elder health care. In literature reviews, there are three main possible approaches which include image-based technique, context aware method, and acceleration-based detection approach.
Nowadays, due to the rapid growing of population using smartphone and the development of hardware technique, the performance of mobile devices is also improved dramatically. Smartphones involve with daily activities of people even elder. In this paper, we proposed a smartphone-based system which utilizes the built-in inertial sensors to analyze the daily motion of human behavior and applies Hidden Markov Model (HMM) to build the model of falls detection. The proposed model system relies on statistical probability of detection to generate the prediction of fall states occurrence.
Generally, the system consists of three main parts: 1) Feature Vector Getting Part, 2) Training Part, and 3) Fall Detection Generating Part. As our experiment, human behavior is a continuity of the movement. Hence, we employed a state transition matrix to illustrate the transition between multiple movement states. All states were connected to result a complete HMM model. We firstly applied Viterbi algorithm to obtain the probability of falls, and then performed necessary processes to achieve detection and prediction purposes. The experiment result shows that the accuracy of fall detection module of our system is up to 93.7% which reduces the large potential risk of elder injuries.

並列關鍵字

Fall Detection ； Hidden Markov Mode ； 3 axis-accelerometer

參考文獻

[1] L. Hui-Tang, H. Yao-Jung, C. Ming-Che, and C. Wang-Rong “ActionView: A Movement-Analysis Ambulatory Monitor in Elderly Homecare Systems” Institute of Computer and Communication Engineering, National Cheng Kung University, pp. 3098-3101, 2009.

[2] Y. Miaou, R. Adel and N. Syed Mohsen “A Posture Recognition-Based Fall Detection System for Monitoring an Elderly Person in a Smart Home” Information Technology in Biomedicine, pp. 1274 – 1286, 2012.

[3] H. Shang-Lin,S. Ming Hsiung,L. Lu Feng and J. Wey-Wen, “A Finite State Machine-Based Fall Detection Mechanism on Smartphones” Ubiquitous Intelligence & Computing and 9th International Conference on Autonomic & Trusted Computing (UIC/ATC), pp. 735-739, 2012.

[4] Yue Shi, Yuanchun Shi and Xia Wang, “Fall Detection on Mobile Phones Using Features from a Five-Phase Model” Ubiquitous Intelligence & Computing and 9th International Conference on Autonomic & Trusted Computing (UIC/ATC), pp. 951-956, 2012.

[5] A. Purwar,J. Do un and C. Wan Young “Activity Monitoring from Real-Time Triaxial Accelerometer data using Sensor Network” International Conference on Control Automation and Systems, pp. 2402-2406, 2007.