Prediction of future action start in untrimmed videos based on diffusion

In recent years, Online Detection of Action Start has attracted considerable interest owing to its diverse applications in areas like healthcare and smart transportation. However, the inherent limitation of detection tasks—"detection can only occur after an event has happened"—prevents us from obtaining responses and issuing warnings before certain events, particularly emergencies, occur. In order to mitigate this constraint, this paper introduces the Prediction of Action Start (PAS) task based on ODAS. The PAS task aims to predict the start times and categories of future actions (if any) in untrimmed streaming videos containing substantial background footage, based on recently observed video information. Our experimental results on two standard datasets demonstrate that the proposed prediction method even outperforms some current action detection methods.