Energy-Based Motion Pattern Detection with Partial Distance Correlation

Abstract

This paper proposes a skeleton-based action classification framework that integrates an Energy-Based Model (EBM) with partial distance correlation (pdCor) to capture nonlinear dependencies among motion features. The method is evaluated under three experimental settings, including unconditional distance correlation analysis, conditional distance correlation analysis, and comparative classification performance assessment. Experimental results demonstrate the presence of significant nonlinear relationships among velocity, positional, and joint state features. By incorporating pdCor, indirect effects among variables are mitigated, enabling more reliable estimation of direct dependencies and improving feature representation quality. The proposed framework achieves an accuracy of 0.71, with a macro-averaged F1-score of 0.77 and a macro-averaged AUC of 0.93, outperforming conventional deep learning and baseline EBM models. These results confirm that integrating EBM with conditional distance correlation enhances discriminative capability between similar motion patterns and improves overall system robustness, indicating strong potential for real-time action recognition and complex motion analysis applications