Fall-down event detection for elderly based on motion history images and deep learning

The goal of this research is to apply the state-of-the-art deep learning approach to human fall-down event detection based on Motion History Images (MHI) from multiple color video sequences captured at different viewing angles. MHI is derived by detecting and combining temporal 2D human contours from surveillance cameras. A human action can then be represented by several continuous MHI images. We then use deep learning approach (CNN + LSTM architectures) to recognize the fall-down behavior from MHI sequences. Our method is capable of not only recognizing the actions of walking, standing, falling down, but also rising after falling down to avoid excessive false alarms. The accuracy of classification into the above 4 short-term actions is capable of achieving 97.66%. We also compare the performances of deep learning architectures that use simple CNN or CNN+LSTM, one or two-stage training, and single or two cameras. Our contributions lie on two aspects: (1) improving the performance on human action recognition based on MHIs and a combination of CNN+LSTM architecture, (2) preventing the false alarm of falling-down events that actually need no help.