- 學位論文
基於Kinect的跌倒偵測與行為監控系統
A Kinect-based Fall Detection and Activity Monitoring System
摘要
由於獨居老人與醫院的病患經常處在跌倒後延遲救助的巨大風險之中，因此近年來有關跌倒偵測系統的研究有蓬勃的發展。本論文的目標是開發一套基於Kinect的跌倒偵測與行為監控系統，替獨居老人提供更安全的生活環境。系統捨棄骨架追蹤資訊，取Kinect的原始深度資訊來做使用。起初，系統會先偵測出原始地面資訊，然後，藉由一個簡單的動態背景相減演算法，從目前所偵測到的地面資訊與原先的地面資訊間的差異中找出前景資訊；當一個具有一定體積與高度的前景物出現，系統會將其視為獨居老人的候選目標，追蹤並分析此移動物體。系統將老人的日常行為用決策樹劃分為五個主要種類：站、走、坐、躺、蹲。當一個跌倒事件被偵測到時，系統會立即發出警示給照顧者。此外，系統會自動生成跌倒事件紀錄 (如：時間、地點、如何跌倒等)，提供更多有意義的資訊給醫療照護者。除了跌倒偵測功能之外，本篇論文所提出的系統亦可提供日常的行為資訊(如：躺在床上的時段、進入廁所的時段、行走的時段等)以供後續分析使用。 實驗的設計上分別設計三個情境，情境一是測試在只有正常活動下是否會有誤判情形發生；情境二是測試是否在沒有環境變動下能正確偵測跌倒事件；情境三是測試是否在有環境的變動之下能正確偵測跌倒事件。在偵測跌倒的實驗中共有90次的跌倒事件，其精確率(Precision)約為94%，召回率(Recall)約為96%。
並列摘要
Older individuals living alone at home or wards of a hospital are usually at great risk of delayed assistance following a fall; therefore, the research of fall detection systems has been greatly growing in these years. This thesis aims to develop a Kinect-based fall detection and activity monitoring system to provide more safety for older individuals living alone. The raw data from the Kinect depth images are processed directly rather than the skeletal tracking information. The system starts from the detection of the ground plane and then a simple dynamic background subtraction algorithm is used to identify foreground pixels from the changes between the currently detected ground plane and the original background ground plane. A foreground object with at least a minimum size and height is considered to be a candidate of the older individual and then this moving object will be tracked and analyzed. Decision trees are adopted to divide the daily activities of the older individual into five major types: standing, walking, sitting, lying, and squatting. The system will issue a warning signal to caregivers whenever a fall event is detected. In addition, the system will automatically generate a fall events record (e.g., when, where and how the fall happened, etc) which provides much valuable information for the health care providers. In addition to the detection of falls, the proposed system can also provide information about the daily activities (e.g., the time period of lying in a bed, time period of entering a lavatory, time period of walking, etc.) for the analysis purpose. The performance of the proposed system was verified by three experimental scenarios. The first experimental scenario is designed to test whether false alarms would happen under normal daily movements. The second experimental scenario is designed to test whether falls could be correctly detected under no change of environments. The third experimental scenario is designed to test whether falls could be correctly detected if there are some changes in environments. Among 90 fall events, the precision ratio and the recall ratio were 94% and 96%, respectively.