Perception of Human Movement Based on Modular Movement Primitives

People can identify and understand human movement from very degraded visual information without effort. A few dots representing the position of the joints are enough to induce a vivid and stable percept of the underlying movement. Due to this ability, the realistic animation of 3D characters req...

Full description

Saved in:
Bibliographic Details
Main Author: Knopp, Benjamin
Contributors: Endres, Dominik (Prof. Dr.) (Thesis advisor)
Format: Dissertation
Language:English
Published: Philipps-Universität Marburg 2021
Subjects:
Online Access:PDF Full Text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:People can identify and understand human movement from very degraded visual information without effort. A few dots representing the position of the joints are enough to induce a vivid and stable percept of the underlying movement. Due to this ability, the realistic animation of 3D characters requires great skill. Studying the constituents of movement that looks natural would not only help these artists, but also bring better understanding of the underlying information processing in the brain. Analogous to the hurdles in animation, the efforts of roboticists reflect the complexity of motion production: controlling the many degrees of freedom of a body requires time-consuming computations. Modularity is one strategy to address this problem: Complex movement can be decomposed into simple primitives. A few primitives can conversely be used to compose a large number of movements. Many types of movement primitives (MPs) have been proposed on different levels of information processing hierarchy in the brain. MPs have mostly been proposed for movement production. Yet, modularity based on primitives might similarly enable robust movement perception. For my thesis, I have conducted perceptual experiments based on the assumption of a shared representation of perception and action based on MPs. The three different types of MPs I have investigated are temporal MPs (TMP), dynamical MPs (DMP), and coupled Gaussian process dynamical models (cGPDM). The MP-models have been trained on natural movements to generate new movements. I then perceptually validated these artificial movements in different psychophysical experiments. In all experiments I used a two-alternative forced choice paradigm, in which human observers were presented a movement based on motion-capturing data, and one generated by an MP-model. They were then asked to chose the movement which they perceived as more natural. In the first experiment I investigated walking movements, and found that, in line with previous results, faithful representation of movement dynamics is more important than good reconstruction of pose. In the second experiment I investigated the role of prediction in perception using reaching movements. Here, I found that perceived naturalness of the predictions is similar to the perceived naturalness of movements itself obtained in the first experiment. I have found that MP models are able to produce movement that looks natural, with the TMP achieving the highest perceptual scores as well highest predictiveness of perceived naturalness among the three model classes, suggesting their suitability for a shared representation of perception and action.
Physical Description:108 Pages
DOI:https://doi.org/10.17192/z2021.0225