A Novel Full-Body and Geometric Features for Physical Sports Interaction Recognition

This paper presents a novel framework for accurately recognizing physical sports interaction within video sequences. The developed approach is based on using image processing and machine learning tools to extract and classify the features correctly. The framework consists of five main steps: preprocessing, silhouette extraction, feature extraction, feature optimization, and classification. Gamma correction enhances image contrast, followed by silhouette extraction using the Multiple Object Tracking (MOT) algorithm and graph-based segmentation. Brief and geometric skeleton-based keypoints are extracted to capture discriminative features, and Independent Component Analysis (ICA) is utilized to optimize feature representation. Finally, the optimized features are classified using the XGBoost algorithm, achieving an impressive accuracy of $\mathbf{9 2 \%}$. The proposed framework demonstrates its effectiveness in recognizing physical sports interaction, paving the way for applications in various domains such as human-human interaction and surveillance.