Human Motion Analysis and Synthesis in Computer Graphics

Shen, Yijun (2019) Human Motion Analysis and Synthesis in Computer Graphics. Doctoral thesis, Northumbria University.

Text (Doctoral Thesis)
shen.yi-jun_phd_13042714.pdf - Submitted Version

Download (3MB) | Preview


This thesis focuses on solving a challenging problem in the field of computer graphics, namely to model and understand 3D human motion efficiently and meaningfully. This is vital to achieve the analysis (health & sports science), synthesis (character animation) and control (video game) of human movements. Though numerous studies have focused on improving the results of motion analysis, motion synthesis and motion control, only a few of these studies solved the problems from the fundamental part owing to the lack of information encoded in motion data.

In my works, the motion of human was divided into the three types, namely single human motion, multi-people interactions and crowd movement. Subsequently, I solved the problems from motion analysis to motion control in different types of motion.

In the single human motion, two types of motion graphs on the motion sequence were proposed using Markov Process. The human motion is represented as the directed graphs, which suggests the number of action patterns and transitions among them. By analyzing the graphs topologies, the richness, transitions flexibility and unpredictability among different action patterns inside the human motion sequence can be easily verified. The framework here is capable of visualizing and analyzing the human motion on the high level of action preference, intention and diversity.

For the two people interaction motion, the use of 3D volumetric meshes on the interacting people was proposed to model their movement and spatial relationship among them. The semantic meanings of the motions were defined by such relationship. A customized Earth Movers Distance was proposed to assess the topological and geometric difference between two groups of meshes. The above assessment captured the semantic similarities among different two-people interactions, which is consistent with what humans perceive. With this interaction motion representation, the multi-people interactions in semantic level can be retrieved and analyzed, and such complex movements can be easily adapted and synthesized with low computational costs.

In the crowd movement, a data-driven gesture-based crowd control system was proposed, in which the control scheme was learned from example gestures provided by different users. The users gestures and corresponding crowd motions, representable to the crowd motions properties and irrelevant to style variations of gestures and crowd motions, were modelled into a compact low dimensional space. With this representation, the proposed framework can take an arbitrary users input gesture and generate appropriate crowd motion in real time.

This thesis shows the advantages of higher-level human motion modelling in different scenarios and solves different challenging tasks of computer graphics. The unified framework summarizes the knowledge to analyze, synthesize and control the movement of human.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: Machine learning , computer animations, Human-Computer Interactions, information retrieval, mathematical modelling
Subjects: G400 Computer Science
G600 Software Engineering
Department: Faculties > Engineering and Environment > Computer and Information Sciences
University Services > Graduate School > Doctor of Philosophy
Depositing User: John Coen
Date Deposited: 18 Mar 2022 09:25
Last Modified: 18 Mar 2022 09:30

Actions (login required)

View Item View Item


Downloads per month over past year

View more statistics