A multi-person motion prediction method and device based on real physical condition constraints

Abstract

The application discloses a multi-person motion prediction method based on real physical condition constraints, comprising the following steps: acquiring a historical action sequence and inputting the historical action sequence into a space-time representation learner to output corresponding space-time representation; inputting the space-time representation into a pre-constructed multi-person motion prediction model to output a corresponding multi-person motion sequence; constructing a loss function about real joint coordinates and predicted joint coordinates based on the historical action sequence and the multi-person motion sequence; constructing a dynamic equilibrium condition based on Euler-Lagrange equation and introducing the dynamic equilibrium condition into the loss function to construct a total loss function, and optimizing the multi-person motion prediction model based on the total loss function to obtain a high-quality multi-person motion prediction model. The application further provides a multi-person motion prediction device. The method provided by the application can effectively eliminate physical abnormal phenomena such as joint penetration and incoherent motion.

Images