Predicting Standing Reach Postures using Deep Neural Networks

Abstract

Proactive ergonomic analysis of occupational tasks using digital human figure models requires accurate prediction of worker postures. A wide range of methods have been proposed and used, including posture libraries, statistical methods including regression, and optimization approaches that incorporate hypothesized criteria such as strength maximization. A common challenge in the implementation of any method is ensuring that the resulting postures are consistent with the kinematic linkage of the figure model, the boundary constraints are satisfied, including those relevant to the task, and that the figure remains in balance after taking into account external forces. Neural network methods have been applied to human posture prediction for more than 25 years, but successful implementation for human posture prediction requires careful consideration of the relevant constraints. This paper describes the implementation of DNN methods within the Human Motion Simulation Framework, which provides a hierarchical structure for posture and motion prediction applicable to any human figure model.