A Reproducible Skeleton Baseline and Multimodal Fusion Protocol for Movement-Pattern Atypicality Analysis Using CZU-MHAD

Abstract

Public multimodal action datasets are primarily designed for action recognition rather than the assessment of within-action movement patterns. This study presents a reproducible skeleton baseline and a specified multimodal fusion protocol using the Changzhou University Multi-modal Human Action Dataset (CZU-MHAD). The empirical analysis processed 357 skeleton recordings, with participant-holdout evaluation conducted on 294 recordings representing 15 shared action categories. Skeleton trajectories were anatomically normalised and evaluated using kinematic logistic regression, principal-component logistic regression, and a temporal convolutional network (TCN). The kinematic baseline achieved mean accuracy of 0.817 and macro-F1 of 0.813 across the two directional participant holdouts, outperforming the TCN, which achieved accuracy of 0.624 ± 0.063 and macro-F1 of 0.601 ± 0.059. Action-conditional embedding distances were additionally used to describe movement-pattern atypicality, without interpreting atypicality as incorrect technique or clinical abnormality. A lightweight multimodal pilot provided preliminary evidence that inertial information can add participant-independent discriminative signal. The study therefore establishes a reproducible empirical baseline while providing a clearly bounded protocol for subsequent depth–skeleton–inertial fusion evaluation.