A machine learning approach for non-invasive fall detection using Kinect

Human falls seldom occur; however, predicting falls is critical in health and safety. When aged people are alone at home, the chances of a dangerous fall are high. There is no one to help them instantly; therefore, timely notifying the concerned caregivers is crucial. Our approach has created a low computational cost algorithm, which depends on the head joint extracted using Kinect. The standard deviation formula is used to check variation in the Y-axis of head joints in every frame. The rate of change is then compared to the value of standard deviation generated from experimental results. After calculating the head trajectory’s standard deviation, the next step is to classify every event into a separate fall and non-fall using the k-NN model. For assessing the k-NN model, we captured around 120 video samples of different activities tested by performance metrics to achieve an accuracy of 0.95.