A computational framework for identifying salient moments in motion capture data

This study presents a computational framework for identifying salient moments for the analysis of motion capture data, with a specific focus on the action of throwing a ball. Identifying these moments is helpful for progress in fields such as animation, sports science, and physical therapy, as it allows for a detailed analysis of complex movements. Traditionally, identifying key moments in motion capture sequences has been based on subjective and inconsistent manual observation. Our study aims to overcome this issue by employing unsupervised techniques, including time series interpolation, differential spatial analysis, spectral clustering, and determination of the optimal number of clusters, to detect important movements within motion capture sequences systematically. We conducted an online study to validate our framework where participants identified key video sequence moments. These were then cross-referenced with motion capture data. The empirical validation showed significantly higher congruence between computationally identified salient moments and those validated by participants with respect to an alternative approach that does not incorporate spectral clustering. The proposed approach offers a standardized, objective method for extracting key moments from motion capture data.