隨著全球人口高齡化與醫療照護人力短缺等現狀，居家健康照護成為當前民生重要議題年長者或獨居者在家中活動，跌倒受傷是普遍存在的風險，尤其是年長者一但跌倒時，若未獲得及時的協助，恐將造成嚴重的傷害。近年來許多跌倒警示系統和穿戴式跌倒警示裝置陸續提出，其中基於相機光學輔助的跌倒事件偵測技術與應用引起廣泛的研究關注，然而，在居家生活環境中這類的跌倒偵測方法面臨不少限制，例如障礙物遮擋及相機視幅和視角等因素。 因此本篇論文提出一套基於深度強化學習之多相機協作跌倒偵測機制，透過多台相機裝置之間進行協作與判斷，來解決單一相機在跌倒事件偵測時所遇到的困難，並且利用深度強化學習的方式來針對多相機協作的動態群組進行學習，目的是為了提升多相機系統的準確性以及加快系統決策的時間，並且在本論文中我們透過實際建置實驗環境和實作出系統之雛形開發，並且以跌倒偵測來作為我們的實施例，之後在針對單相機決策、多相機決策（未使用動態群組）以及多相機決策（使用動態群組）三種方案來進行實際的效能比較。 ;With the aging of the global population and the shortage of medical care manpower, home health care has become an important part of people’s livelihood. To issue. If the elderly or those living alone are active at home, falls and injuries are a common risk, especially for the elderly. However, if you do not get timely assistance when you fall, you may cause serious injuries. In recent years, many fall warning systems and Wearable fall warning devices have been proposed one after another. Among them, the fall event detection technology and application based on the optical assist of the camera are introduced. From a wide range of research concerns. However, such fall detection methods face many limitations in the home living environment, such as Factors such as occlusion by obstacles and the camera′s field of view and angle of view.
Therefore, this paper proposes a multi-camera collaborative fall detection mechanism based on deep reinforcement learning. Image recognition collaboration and judgment between multiple cameras to solve the difficulties encountered by a single camera in the fall event detection Difficult, and use deep reinforcement learning to learn for dynamic groups of multi-camera collaboration, the purpose It is to improve the accuracy of the multi-camera system and speed up the decision-making time of the system, and through the actual construction Detection environment for single-camera decision-making, multi-camera decision-making (not using dynamic groups), and multi-camera decision-making (Using dynamic groups) Three schemes for actual performance comparison.
