Fall Detection Using Angle-Based Feature Extraction from Human Skeleton and Machine Learning Approach

Mitigating fall events is a significant challenge in modern societies, particularly in the context of enhancing the well-being and assisting in the daily activities of older individuals. As a consequence, the annual incidence of falls among the elderly is on the rise, and the potential peril becomes more significant if identification and intervention occur too tardily. To observe and notify fall situations in real time, many researchers have been working to develop a fall detection system using various sensor information. Though their system, in some cases, generated high detection rates, many older people were reluctant to carry the sensor, which is a barrier to normalizing the systems for common use. In addition, the vision-based system still faces various difficulties, such as redundant background, light illumination and partial occlusion, to achieve high efficiency and performance of the model. We proposed a skeleton-based fall detection approach using angle features and a machine learning approach to overcome the problem. In the procedure, we first extracted the whole-body landmark key points using a media pipe and then we calculated the 70-angle features and employed the Boruta feature selection approach to select the potential features. Finally, a support vector machine (SVM) with various hyperparameter tuning is used for the classification. The evaluation of our proposed model was conducted using the UR fall detection video dataset. The outstanding performance observed in the results highlights the superiority of our method over the current state-of-the-art approaches.