Sequential Unsupervised Motion Segmentation for Skill Evaluation and Comparison

Ken Kajiwara, Issei Saito, Tomoaki Nakamura, Daichi Mochihashi, Koki Mimura

Abstract: In recent years, the evaluation of motion skills in fields such as dance and exercise has benefited from the proliferation of motion-capture technologies. This study proposes a computationally efficient method for segmenting and comparing time-series human motion data across individuals with varying skill levels. The method extends the Gaussian process hidden semi-Markov model (GP-HSMM) by incorporating random Fourier features (RFF), enabling motion segments to be modeled through linear regression and significantly reducing computational complexity. We also introduce sequential learning to support the incremental adaptation of the model to new individuals, which facilitates quantitative motion comparisons on a per-segment and per-joint basis. To preserve segmentation accuracy despite the approximation introduced by RFF, we optimize the kernel parameters using the Metropolis-Hastings algorithm within a Markov chain Monte Carlo (MCMC) framework. Experiments on publicly available exercise and dance datasets show that the proposed method achieves segmentation performance similar to GP-HSMM while reducing training time from several days to minutes. Moreover, the method supports the quantitative visualization of motion-skill differences, illustrated through a joint-level analysis of choreographic elements performed by dancers with different skill levels.