近年來社會面臨老年化日趨嚴重的問題，老人照護議題也越發重要；老人跌倒的發生頻率高，且常附帶著因延後就醫所帶來的巨大風險，因此跌倒偵測的相關研究越來越蓬勃發展。本論文開發一套基於結合Kinect 及類神經網路的跌倒偵測系統，希望在多人情境下，能夠正常運作，使其更貼近現實生活。 本系統使用Kinect 所提供之深度資訊，找出場景中的地面資訊，再搭配背景相減演算法，找出前景資訊，並進一步追蹤；再利用事前訓練好的類神經網路模型及選定的特徵，判斷是否有跌倒事件發生，當系統偵測跌倒時，會記錄當下的畫面以及時間，回傳給照顧者。本論文會針對規則式判斷可能會發生誤判的情況做探討，及利用類神經網路所帶來的優勢。 本論文設計了六種模擬情境，三種單人情境、三種多人情境，並且比較規則式判斷及使用類神經網路的實驗結果，探討之間的差別以及誤判的可能因素。在全部的情境中總共有168 次跌倒事件以及168 次未跌倒事件，其實驗結果，敏感度(Sensitivity)約為97%，特異度(Specificity)約為90%，Kappa 值為0.84，證明系統有幾乎與事實吻合的程度。
A Fall Detect System Based on Neural Networks with Kinect Depth-Camera
Recently, society is faced with the problematic issue of an aging population. The eldercare issue is extremely important. The frequency of falls in the elderly is higher than in younger people with a greater risk caused by treatment delay. Therefore, the research of fall detection systems has been increasing drastically. This thesis proposes to develop a fall detection system based on neural networks with Kinect depth-camera. We hope it can operate reliable in a complex environment or in multi-person scenarios. The system uses raw data of Kinect depth images to locate the ground in the scene, identify the foreground pixels with a background subtraction algorithm, and then tracked the foreground for analysis. Last, the system will judge whether the fall events occurred by using its well-trained neural networks model and the specified features. When fall events are detected, the system would record the image and time immediately, and then report to caregivers for efficient aid. Additionally, this thesis will discuss the reasons for rule decision system's misjudgment and the advantages of using neural networks. The performance of the proposed system was verified by six experimental scenarios. There are three single person and for multi-person experimental scenarios. After these experiments, we would compare the result of rule decision system with the proposed system and discuss the difference and the reason of misjudgment between both of them. Among all of these experimental scenarios: 168 are fall events and 168 are not fall events. The results show the sensitivity iii rate and the specificity rate were 97% and 90%, respectively. And the Kappa value of the proposed system is 0.84 which is higher than 0.80, showing that we have a reliable system that accurately reflects reality in terms of fall events.
Keywords : Kinect, fall detect system, video surveillance, eldercare, neural networks.