I see what you do: The HumAct algorithm for live human action recognition and applications

Realtime human action recognition from video streams is becoming increasingly essential for applications that require immediate interpretation of human behavior. This includes scenarios such as interactive virtual environments and responsive healthcare systems. This paper presents HumAct, a novel algorithm developed for live human action recognition. It reliably detects a broad range of human actions, even within diverse and challenging video content. The algorithm achieves realtime performance, identifying actions for multiple persons in the scene with a delay of only two seconds from their occurrence in a video live stream. We demonstrate the practical effectiveness of the HumAct algorithm through two applications: enhancing intuitive interaction in VR/XR environments and delivering timely alerts for falling elderly persons in an ambient assisted living setting. Together, these examples highlight the flexibility and strong potential of our proposed live human action recognition algorithm for real-world deployment.