Investigating similarity measures for locomotor trajectories based on the human perception of differences in motions

Abstract

Providing robots with the ability to move humanlike is one of the recent challenges for researchers who work on motion planning in human populated environments. Human-like motions help a human interaction partner to intuitively grasp the intention of the robot. However, the problem of validating the degree of human-likeness of a robot motion is rarely addressed, especially for the forward motion during navigation. One approach is using similarity measures to compare the robot trajectories directly with human ones. For this reason, this paper investigates different methods from the time series analysis that can be applied to measure the similarity between trajectories: the average Euclidean distance, the Dynamic Time Warping distance, and the Longest Common Subsequence. We aim to identify the measure that performs the same way as a human who rates the similarity. Thus, the evaluation of the methods is based on a questionnaire that examines the human perception of differences between walking motions. It is concluded that the human similarity perception is reproduced best by using the Dynamic Time Warping and comparing the derivatives of the path and velocity profiles instead of the absolute values.