Human Fall Detection using Skeleton Features

Unintentional falls of people, when left without serve may cause severe injuries and in extreme cases, they may even lead to loss of lives. In order to provide timely medication, detection of fall events when occurred is necessary.Any action can be considered as the specific motion of various bone key points. So, in our work, we considered bone key points as feature extractors. The MediaPipe framework developed by Google is used to detect the bone key features and its coordinates on the human skeleton. The data obtained is then normalized with respect to the boundary box drawn around humans. Machine learning classifiers, RF, SVM and Deep Learning model, DNN are then used individually to recognise and classify the action into fall or non-fall events. NTU-RGB+D dataset is used in our work. Real time detection using a webcam is also tested. The accuracy achieved by DNN model is 97.63% and that of SVM and RF classifiers is 83.3% and 99.34% respectively. Thus, the highest accuracy is achieved by RF classifier which is 99.34%.