On-line Time Warping of Human Motion Sequences

Randall, Mathew (2023) On-line Time Warping of Human Motion Sequences. Doctoral thesis, Birmingham City University.

[img]
Preview
Text
Mathew Randall PhD Thesis published_Final version_Submitted May 2023_Final Award Oct 2023 .pdf - Accepted Version

Download (39MB)

Abstract

Some application areas require motions to be time warped on-line as a motion is captured, aligning a partially captured motion to a complete prerecorded motion. For example movement training applications for dance and medical procedures, require on-line time warping for analysing and visually feeding back the accuracy of human motions as they are being performed. Additionally, real-time production techniques such as virtual production, in camera visual effects and the use of avatars in live stage performances, require on-line time warping to align virtual character performances to a live performer.

The work in this thesis first addresses a research gap in the measurement of the alignment of two motions, proposing approaches based on rank correlation and evaluating them against existing distance based approaches to measuring motion similarity. The thesis then goes onto propose and evaluate novel methods for on-line time warping, which plot alignments in a forward direction and utilise forecasting and local continuity constraint techniques.

Current studies into measuring the similarity of motions focus on distance based metrics for measuring the similarity of the motions to support motion recognition applications, leaving a research gap regarding the effectiveness of similarity metrics bases on correlation and the optimal metrics for measuring the alignment of two motions. This thesis addresses this research gap by comparing the performance of variety of similarity metrics based on distance and correlation, including novel combinations of joint parameterisation and correlation methods. The ability of each metric to measure both the similarity and alignment of two motions is independently assessed.

This work provides a detailed evaluation of a variety of different approaches to using correlation within a similarity metric, testing their performance to determine which approach is optimal and comparing their performance against established distance based metrics. The results show that a correlation based metric, in which joints are parameterised using displacement vectors and correlation is measured using Kendall Tau rank correlation, is the optimal approach for measuring the alignment between two motions. The study also showed that similarity metrics based on correlation are better at measuring the alignment of two motions, which is important in motion blending and style transfer applications as well as evaluating the performance of time warping algorithms. It also showed that metrics based on distance are better at measuring the similarity of two motions, which is more relevant to motion recognition and classification applications.

A number of approaches to on-line time warping have been proposed within existing research, that are based on plotting an alignment path backwards from a selected end-point within the complete motion. While these approaches work for discrete applications, such as recognising a motion, their lack of monotonic constraint between alignment of each frame, means these approaches do not support applications that require an alignment to be maintained continuously over a number of frames. For example applications involving continuous real-time visualisation, feedback or interaction.

To solve this problem, a number of novel on-line time warping algorithms, based on forward plotting, motion forecasting and local continuity constraints are proposed and evaluated by applying them to human motions. Two benchmarks standards for evaluating the performance of on-line time warping algorithms are established, based on UTW time warping and compering the resulting alignment path with that produced by DTW. This work also proposes a novel approach to adapting existing local continuity constraints to a forward plotting approach.

The studies within this thesis demonstrates that these time warping approaches are able to produce alignments of sufficient quality to support applications that require an alignment to be maintained continuously. The on-line time warping algorithms proposed in this study can align a previously recorded motion to a user in real-time, as they are performing the same action or an opposing action recorded at the same time as the motion being align. This solution has a variety of potential application areas including: visualisation applications, such as aligning a motion to a live performer to facilitate in camera visual effects or a live stage performance with a virtual avatar; motion feedback applications such as dance training or medical rehabilitation; and interaction applications such as working with Cobots.

Item Type: Thesis (Doctoral)
Dates:
DateEvent
26 May 2023Submitted
3 October 2023Accepted
Uncontrolled Keywords: Motion capture, human motion, motion similarity, motion alignment, time warping, on-line time warping
Subjects: CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 20 Oct 2023 11:29
Last Modified: 20 Oct 2023 11:29
URI: https://www.open-access.bcu.ac.uk/id/eprint/14852

Actions (login required)

View Item View Item

Research

In this section...