Predicting Human Motion using the Unscented Kalman Filter for Safe and Efficient Human-Robot Collaboration

Abstract

Predicting human motion is vital for enhancing safety and efficiency in human-robot collaboration. Researchers have dedicated significant efforts to developing accurate human models, often involving optimization and task-specific information. However, regardless of complexity, all models come with uncertainties that robots need to recognize to make informed decisions. This paper examines the performance of two simple models using the Unscented Kalman Filter (UKF) to filter and predict future human poses. Moreover, a combined version of the models is implemented using an Interacting Multiple Model (IMM) estimator. The objective is to evaluate the algorithms' prediction accuracy and uncertainty across various human-robot interaction scenarios under different operating conditions. This analysis identifies suitable settings where the simple model can be effective and highlights situations where a more complex system might be necessary.