RT-ADI: Fast Real-Time Video Representation for Multi-view Human Fall Detection

Falling accidents are potential risks that significantly threaten the elders' health, which leads to severe injuries, such as hip fractures, broken bones or head injury. Nowadays, modern surveillance technologies have been adopted to prevent elders from falling. In this paper, we propose a multi-view human fall detection system with a fast video representation method called RT-ADI to extract video temporal features with low time consumption. We also design a multi-branch convolutional neural network to merge computed dynamic images from multiple cameras installed. Furthermore, we have evaluated our approach on Multiple camera fall dataset. Our system exhibits low time consumption and high accuracy with almost no miss-classification.