University of Glasgow
Learning low-dimensional latent spaces for Hand-pose interaction systems
Abstract
dc:description.abstractIn recent years, mid-air gesture recognition and control have emerged as touch-free interaction mechanisms with significant potential, but human-centered design challenges persist, particularly due to the complexity of user hand movements, user-learnability of interaction system, and usability across varied contexts. This research addresses limitations in current mid-air gesture systems by introducing strategies that reduce the latency of real-time interaction and enhance hand pose interpretability. We introduce a hand pose recognition approach using only two video frames, optimizing for speed and accuracy while minimizing time dependencies, thereby allowing users to experience a more responsive system. Building on these insights, the Hand-pose Embedding Interactive System (HpEIS) employs a Variational Autoencoder to map gestures to a two-dimensional space, introducing visualized feedback mechanisms that improve user experience. At the same time, interaction stability improved through smoothing and anti-jitter methods. While this approach improves robustness in dynamic movements, further challenges remain in adaptability and flexible user control. To expand flexibility, the HandSolo model introduces a disentangled hand pose embedding space, supporting multi-dimensional control with independent degrees of freedom, thus enabling interactions adaptable across devices and contexts. Coupled with a Visual Interaction Evaluation Strategy (VIEs), HandSolo provides a guidance for system designers to align model capabilities with user preferences. The experimentation underscores the effectiveness of these systems. This integrated research establishes a framework for mid-air hand pose control, advancing usability, flexibility and extensibility in the interaction design of mid-air hand pose with high dimensional input.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Xu, Songpei
Subjects
dc:subject × 3Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.85455