由於社會人口結構不斷老化，跌倒成為老年人慢性致殘的第三大原因，因此照護產業日漸受到重視，跌倒偵測的技術研究也不斷向前邁進。前人研究中分別使用了感應偵測器與影像偵測法，來做跌倒偵測，直到2006年出現使用機器學習法來做跌倒偵測研究。 近年機器學習重新受到重視，主要來自於深度學習神經網路的貢獻，深度神經網路由人工類神經網路發展而來，其神經網路結構更加複雜多變，足以解決許多電腦視覺上的難題，因此本論文提出基於深度學習方式的跌倒偵測系統。核心技術，使用光流影像處理與3D卷積神經網路擷取跌倒影像特徵，再利用訓練完成的LSTM遞迴神經網路做跌倒預測分類。 為了加強本系統照護功能，這裡設計跌倒偵測系統具有後端通報功能，搭建以Infrastructure as a Service(IaaS)為基礎的雲端服務，雲端服務可同時照護多處住家，當前端偵測到跌倒事件，即可通報雲端服務，並協助醫護人員快速處理。 本論文參考前人研究，使用3D CNN與LSTM設計深度神經網路，並使用光流前處理進行改良，經實驗後Accuracy由68.47%提升至93.18%；另設計不同分類器實驗 (SVM、MLP)與本跌倒偵測系統進行比較，實驗結果都證實，本論文提出的方法擁有較佳的表現水準。 ;Since falls are a major public health problem among older people, the number of systems aimed at detecting them has increased dramatically over recent years. In previous studies, sensor detectors and image detection methods were used in fall detection research, until 2006, some research used machine learning to do fall detection. In recent years, machine learning has been paid more attention, mainly from the contribution of deep learning. Deep learning is developed by artificial neural network, and its structure is more complex and varied, enough to solve many computer vision problems. This thesis proposes a fall detection system based on the deep learning. The core technology utilizes the optical image processing and 3D convolution neural network to extract features of images, and then to do fall prediction by training LSTM recurrent neural network. In order to strengthen the system function of health-care, the front-end fall detection system has a back-end notification function design. To build a cloud service based on Infrastructure as a Service (IaaS), and the service can take care of multiple homes at the same time. When the front-end detects a fall event, it can notify the cloud service and help the healthcare provider handle it quickly. Based on previous studies, this thesis uses 3D CNN and LSTM to design the deep neural network, and adds the optical flow pretreatment to improve. After the experiment, the accuracy increased from 68.47% to 93.18%. Here also designs different classifiers (SVM, MLP) experiments and compare with the proposed fall detection system. The experimental results confirm that the proposed method has a better performance level.
