To address the problems caused by current measures to improve the life quality of the elderly, such as the cumbersome wearing of sensors and excessive resource acquisition requirements, we designed and implemented an embedded fall-detection system, utilizing MediaPipe technology to estimate and recognize body posture and movement. The recognized feature sequence information, including the X, Y, and Z coordinates, is imported into a long short-term memory (LSTM) neural network for classification, so as to achieve fast detection of falling action. In experiments on a public dataset, Le2i and Multiple cameras fall dataset, the proposed fall-detection system achieved 98% accuracy, with a high detection speed in real-time video streaming.