Zero-shot Gait Classification with Diffusion Models

Xiaodong Guan, Robert James Gray, Ashwani Jha, Parashkev Nachev

Keywords: Gait Analysis, Generative Models, Diffusion Models

TL;DR: A zero-shot, diffusion-based method for gait assessment from video-derived pose data, detecting subtle abnormalities without supervised training while providing a scalable, objective, and interpretable alternative to existing methods.

Abstract: Movement disorders such as Parkinson’s disease are characterised by complex abnormalities of body motion that resist precise, replicable, and scalable quantification. Subjective clinical scores--the established standard--are limited in expressivity and vulnerable to intra-observer variation; wearable sensor-based methods offer objectivity but with limited anatomical sampling. Remote video-based approaches could deliver both highly expressive and objective quantification of motion, but sufficient labelled samples are hard to obtain under clinical data regimes. Here we develop a diffusion model-based, zero-shot, and human-interpretable approach to gait assessment from video-derived pose data and evaluate it in Parkinson's Disease. Capable of detecting subtle changes in body motion without explicit training, it shows potential for an accurate, robust, and scalable solution, addressing the major limitations of existing methods.

Submission Number: 71