FriendInNeed

We use unique ML algorithm with sensors data to decide if fall happen.
Application collects sensors (accelerometer and gyroscope) data and checks if acceleration reached fall threshold. Then application prepare sensor data for last two seconds (max duration of fall) and sends request to backend. On he backend we have Flask python service with TensorFlow for Machine Learning. Currently we use Deep Neural Network Classifier which gives us accuracy of approximately 80%. So the backend side can decide in more precise and accurate way if the person actually fell to the ground and needs help. Application gets response and now knows it should notify list of predefined and chosen friends by alarm message containing GPS coordinates and call for help. We also have some inactivity period, after application got response from server and before it actually sends alarm messages - it notifies user fall detected and shows timer counter which could be cancelled with button or any significant move. In both cases we also redeliver sensor data to server but this time in labeled format (because user marks it as false positive in case she cancels countdown timer) - and this way we collect data to retrain network on daily basis.