Deep Learning-Based Pose Estimation and Real-Time Toddler Fall Detection System

Many falls happen each year which cause severe medical issues. An early detection could allow the injured toddler to receive medical care right away. Deep learning techniques like long short-term memory (LSTM) are proving to be more efficient than the use of sensors to detect the fall. These new techniques can right away detect the fall and inform the concerned ones in time. We employed a camera-based system along with human pose estimation by using OpenPifPaf to extract the keypoints (elbows, shoulders, knees, etc.) of the toddler, these keypoints are passed on to the LSTM model for classification. The LSTM-based neural network predicts the falls into four classes: ‘Normal’, ‘Fall Warning’, ‘No Fall’, and ‘Fall’. To train this deep learning model, we used the UP-Fall Detection dataset, which consists of videos of various activities. This paper aims to reduce the number of fatal injuries in toddlers by using a deep learning real-time fall detection system that will notify the parents and caregivers as soon as a fall occurs so that the toddler can receive medical care immediately if necessary.