A comparison between optimization-based human motion prediction methods: data-based, knowledge-based and hybrid approaches

Abstract

In this paper an optimization-based hybrid dynamic motion prediction method is presented. The method is hybrid as the prediction relies both on actually performed motions for reference (following a data-based approach) and on the definition of appropriate performance measures (following a knowledge-based approach). The prediction is carried out through the definition of a constrained non-linear optimization problem, in which the objective function is composed of a weighted combination of data-based and knowledge-based contributions. The weights of each contribution are varied in order to generate a battery of hybrid predictions, which range from purely data-based to purely knowledge-based. The results of the predictions are analyzed and compared against actually performed motions both qualitatively and quantitatively, using a measure of realism defined as the distance of the predicted motions from the mean of the actually performed motions. The method is applied to clutch pedal depression motions and the comparison between the different approaches favors the hybrid solution, which seems to combine the strengths of both data- and knowledge-based approaches, enhancing the realism of the predicted motion.