Neural network for dynamic human motion prediction

Abstract

Digital human models (DHMs) are critical for improved designs, injury prevention, and a better understanding of human behavior. Although many capabilities in the field are maturing, there are still opportunities for improvement, especially in motion prediction. Thus, this work investigates the use of an artificial neural network (ANN), specifically a general regression neural network (GRNN), to provide real-time computation of DHM motion prediction, where the underlying optimization problems are large and computationally complex. In initial experimentation, a GRNN is used successfully to simulate walking and jumping on a box while using physics-based human simulations as training data. Compared to direct computational simulations of dynamic motion, use of GRNN reduces the calculation time for each predicted motion from 1-40 min to a fraction of a second with no noticeable reduction in accuracy. This work lays the foundation for studying the effects of changes to training regiments on human performance.