A low-complexity LSTM model for real-time fall detection by a video camera
Kuwae Toshiki, Vasily G. Moshnyaga · 2025
Falls are one of the main causes of injuries among elders. Traditional camera-based fall detection methods require large computational resources and complex video processing algorithms for the automatic detection of falling from highresolution videos. Because such resources are mainly available on cloud servers and induce delays, detecting falls in realtime becomes quite challenging. In this article, we introduce a low-complexity LSTM (Long Short-Term Memory)-based approach for camera-based fall detection in real-time. Unlike related research that uses LSTM for fall recognition, we preprocess each video frame to extract the body key points of a person with MediaPipe and then input them onto the LSTM, trained to identify falls from the time-sequence of key-point coordinates. Experiments show that this formulation considerably lowers LSTM complexity while maintaining the Recall score of 98.8%, Precision of 98.8%, and F1 score of 99.5% for real-time (20fps) fall detection.