近幾十年來,隨著人口老齡化,台灣已進入高齡社會。根據衛生福利部資料指出,跌倒意外已成為長者意外傷害死亡的第二大原因,且跌傷後就醫比例僅8%,長者發生跌倒意外事件已成為不可忽視的議題。因此為了改善低就醫比例之問題,本論文提出一種監控式影像居家跌倒偵測系統,其透過重新訓練後之物件辨識模型偵測畫面中人物,之後進行動作判斷,並且依照判斷的方式分為二種演算法,分別為使用SVM進行動作特徵分類的SVM跌倒偵測演算法(SVMFDA),以及使用深度學習法的SlowFast跌倒偵測演算法(SFFDA),並且二種演算法皆可辨識出除跌倒外其他四種日常行為動作。經由多種跌倒資料集之實驗結果,此二種跌倒偵測方法皆能成功辨識出長者的跌倒事件,其正確率分別達到93% 與95%。
In recent decades, with the aging of the population, Taiwan has entered an aged society. According to information from the Ministry of Health and Welfare, fall accidents have become the second leading cause of death from accidental injuries in the elderly, and only 8% of them seek medical attention after falling. Fall accidents in the elderly have become an issue that cannot be ignored. Therefore, in order to improve the problem of low seek medical attention rate. In the thesis, we proposed a home fall detection system base on video surveillance, detect people by the retrained object detection model, then make action judgments, and divides it into two algorithms according to the judgment method, namely, the SVM fall detection algorithm (SVMFDA) that uses SVM to classify action features, and the use of deep learning Method of SlowFast Fall Detection Algorithm (SFFDA),And both algorithms can recognize five actions including fall. Based on the experimental results of various fall data sets, both two algorithms can successfully identify the fall events of the elderly, with the accuracy of 93% and 95% respectively.