Fall Detection with Neural Networks

Falls in adults over 60 years of age can cause traumas that may require hospitalization of the patient, resulting in temporary disability and even death. In this work, an artificial intelligence system was developed using neural networks to detect the fall of a person and alert a family member or caregiver of the occurrence of this event. A bracelet with a motion sensor, which combines the measurements of a gyroscope, a geomagnetic sensor, and a 3-axis accelerometer, to provide an orientation vector is placed on the user's wrist, allowing the detection and generation of records of their activities. In addition, a Raspberry Pi card was used for data concentration and subsequent processing. The recorded data were evaluated with three neural networks to determine whether they correspond to a “fall” event and at the same time send an alert message via Telegram to family members or caregivers of the elderly. In this study, the operation of three artificial neural networks was compared to classify a person's activity as “falling” or “not falling”. The results show that the convolutional neural network confers higher accuracy (98.5%) and lower loss (0.23), with alert messages arriving instantaneously when the event occurs. Another advantage of this system is that it allows the connection and collection of data from 20 simultaneous devices, to monitor other nearby users with excellent accuracy.