Behavioral Biometric-Based Human Identification from 3D Skeletal Motion Using Convolutional Neural Networks

Accurate and reliable human identification is needed to provide personalized services in conjunction with current and future technological developments. This study presents a motion-based human identification approach that utilizes 3D skeleton data to capture unique individual motion patterns. Joint coordinates are processed to extract spatial and temporal features through two Convolutional Neural Network (CNN) models with different depths. Evaluation on the UTKinect-Action3D dataset (20 joints, 10 subjects, nine actions) shows that the deeper model achieves 98.25% accuracy, whereas the lighter model achieves 96.97%. To assess generalization, both models are further tested on the Florence 3D Actions dataset with 15 joints, achieving accuracies of 85.53% and 83.55%, respectively. These findings confirm that detailed motion representation significantly improves identification performance and demonstrate that CNN-based models can effectively recognize individuals based on body motion patterns.