應用人體姿態估計之跌倒偵測

Abstract

隨著臺灣人口結構中高齡占比持續攀升，跌倒已儼然成為造成老人受傷與死亡的主要因素之一，面對步入高齡化的社會，對於老人醫療照護的需求，未來必然持續增加，但人手不足一直是老人長期照護議題中迫切面臨的問題。為了減輕人力資源短缺的問題，本研究提出一套可於居家環境即時運作的影像式跌倒偵測系統。首先，本研究針對人物偵測模型之跌倒姿態進行重新訓練，辨識出目標人物後再對其進行姿態估計，提取人物關鍵點位置作為輸入特徵，並使用支援向量機(SVM)與多層全連接神經網路進行動作分類，其中全連接神經網路分別使用兩種不同方式的輸入(FCNS, FCNT)。實驗結果顯示，SVM與兩種全連接神經網路皆能辨識包含跌倒動作在內的四種日常生活動作，其平均準確率分別為96.9%、96%與95.6%，且辨識速度能達到實時。此外，本研究設計一套判針對跌倒事件之判斷與通報流程系統也能有效地區分跌倒事件，並分別在URfall資料集上取得了100%、96.6%與100%的準確度。

As Taiwan’s population continues to age, falls have become one of the leading causes of disability and mortality among the elderly. While the demand for geriatric medical care is steadily increasing, chronic staff shortages remain a pressing challenge in long-term care. In order to improve this manpower gap, this thesis proposes a vision-based fall-detection system capable of real-time operation in home environments. First, a person-detection model is re-trained with fall-specific data to localize the target individual. The detected person is then subjected to pose estimation, from which keypoints coordinates are extracted as input features. These features are classified using a Support Vector Machine (SVM) and fully connected neural networks (FCNs), with the FCNs implemented using two different input arrangements. Experimental results demonstrate that the SVM and both fully connected networks are capable of recognizing four daily activities, including falls, with average accuracies of 96.9%, 96%, and 95.6%, respectively. All models achieved real-time inference speeds. In addition, a fall judgment and notification procedure was designed to reduce false alarms. The proposed system successfully distinguished true fall events, achieving accuracy rates of 100%, 96.6%, and 100% on the URfall dataset using SVM, FCNS, and FCNT, respectively.