Fall Detection with a Single Commodity RGB Camera Based-on 2D Pose Estimation

Abstract

In older adults, falls can cause severe injuries. Even right before dropping, reaction time is important, so the near-by device can respond on time. This article introduces a method for early detection of falls that is non-invasive. Our method only uses a single RGB Commodity Camera. With CNN-based, we estimate human-position, then use the pose to recognize falls without any calibration and additional equipment, processing join data and predict with LSTM model. The model achieves 97% in accuracy. It can be extended to a smart home.