針對於長照機構之人體姿態識別及其應用

Abstract

本論文的主要目標為利用彩色影像，結合即時又快速的演算法來進行人體姿態識別佈署在醫院或長照中心以銜接智慧長照場景之各種應用。本研究結合人體估測演算法與合併式模糊類神經網路，提出了一種新的架構來準確地識別人體姿態。首先，我們利用DWPose來從影像中提取人體關鍵點，接著將這些關鍵點進行擴增與不同組合之合併，接著，這些關鍵點被送入合併式模糊類神經網路中進行訓練。針對輸入在不同的擴增與合併組合下，分析出最高準確率的組合，在不同的場景驗證此組合的有效性。根據實驗結果指出所提出方法具有小資料量訓練、受環境影響低、運算速度快的優勢。最後，本文基於此姿態識別，延伸出一些符合長照場景之應用，如跌倒偵測、離房偵測、廁所久待等實際場景。

The main objective of this thesis is to utilize color images combined with real-time and efficient algorithms for human posture recognition, enabling deployment in hospitals or long-term care centers to support various applications in smart long-term care scenarios. This study integrates human pose estimation algorithms with a Merged Fuzzy Neural Network to propose a novel approach for accurately recognizing human postures. Firstly, DWPose is used to extract the human body's keypoints from images, which are augmented and merged in various onfigurations. Subsequently, the keypoints are fed into the Merged Fuzzy Neural Network for training. Different augmentation and merging configurations are analyzed to identify the one with the highest accuracy, which is further validated for effectiveness in various scenarios. Experimental results indicate that the proposed method offers advantages such as low data requirements for training, robustness to environmental influences, and high computational efficiency. Finally, based on this posture recognition, several applications relevant to long-term care scenarios are developed, such as fall detection, room exit monitoring, and prolonged bathroom stay detection in real-life settings.