Fall Detection System Based on YOLO Algorithm and MobileNetV2 Model

Accidental falls are the major causes of injury among elderly people living alone. However, traditional fall detection algorithms usually have missed or false detections problems. In this paper, we design a fall detection system based on YOLO algorithm and MobileNetV2 model. First, detection frame coordinates and joint coordinates of the human body are fused as feature information, and background information is filtered out. A lightweight deep learning model, MobileNetV2, is then used for classification. Compared with commonly used lightweight classification models such as ShuffleNet and EfficientNet, the MobileNetV2 model performs better in recognizing fall behavior. Our system can effectively distinguish fall behavior from daily activities, meeting the requirements for fall detection.