A Real-Time Fall Detection System Based on MoveNet and LSTM

Preventing injuries and deaths caused by the lack of timely assistance to the elderly after a fall is an important issue facing our society. Elderly people are prone to accidents such as falls due to the decline in their physical functions, and once they have fallen, if they do not receive timely assistance, it may lead to more serious consequences and even threaten their lives. Therefore, this study explores a deep learning-based method for fall detection. The objective is to minimize the consequences of such incidents. The study compares 2D pose estimation algorithms and proposes a fall detection algorithm based on MoveNet and Long short-term memory networks. This algorithm utilizes a camera to collect data and extracts skeletal information from videos or images. Long short-term memory is then used to analyze and detect the occurrence of a fall in elderly people. In addition, this study comparatively analyses four 2D pose estimation algorithms, namely OpenPose, BlazePose, PoseNet and MoveNet. The experimental results demonstrate that the MoveNet algorithm exhibits superior real-time performance and accuracy. Through experiments conducted on the UR Fall Detection Dataset, the algorithm achieved high performance. Additionally, the detection frame rate on the CPU reaches up to 33 fps, providing remarkable real-time performance in comparison to other algorithms.