A Fall Detection using Sound Technology Based on TinyML

With the increasing ageing of our society, the need of fall detection for the elderly is becoming more and more important. In this paper, a convolutional neural network model is established and trained to study the fall detection using sound technology, so that the model can realize the autonomous recognition ability of the elderly fall sound from the other noises. This network model will be finally deployed and realized on a small wearable device using TinyML technology. A certain amount of falling sound data and noise data is collected by ourselves, and used as a training data set for neural network. The audio data is converted into a Mel spectrogram firstly, and data features are extracted, and then inputted into the convolutional neural network. This network is trained with TensorFlow offline before deployment. Finally, under an experiment environment, the accuracy of the model's recognition of falling sounds reached more than 95%. Combining the sound technology with acceleration and gyroscope sensors will be a good way to detect the fall of elderly.