2026 Volume 38 Issue 3 Pages 683-689
In human motion analysis, variability across participants and trials, as well as observational noise, are significant factors that impede the accurate extraction of motion features. Traditional techniques, such as simple time normalization or the direct application of Dynamic Time Warping (DTW) in the observation space, struggle to effectively handle these uncertainties, leading to reduced analysis accuracy. To mitigate these challenges, we present a robust framework that applies DTW to trajectories in a low-dimensional latent space produced by a Gaussian Process Dynamical Model (GPDM). By leveraging the GPDM’s probabilistic framework and mean-prediction capability, it becomes possible to extract and synchronize essential motion dynamics in the latent space while eliminating the effects of trial-to-trial variance and noise. Validation through lifting tasks demonstrated that the proposed method preserves the original motion trajectories and achieves stable temporal alignment even under conditions with intense additive noise, effectively suppressing the matching disruptions typically observed when applying DTW in the observation space. Joint-angle synergy analysis of aligned trajectories further confirmed that essential synergies can be consistently extracted, regardless of trial uncertainties.
