Identifying underlying individuality across running, walking, and handwriting patterns with conditional...

Area of Science:

• Biomechanics
• Human movement analysis
• Machine learning

Background:

• Strong individuality in human movement patterns is now accepted.
• Previous studies primarily compared signals from identical movement types.
• A methodological gap exists in analyzing commonalities across different movement patterns.

Purpose of the Study:

• To detect cross-movement commonalities in individual walking, running, and handwriting patterns.
• To utilize data augmentation techniques for this cross-movement analysis.
• To explore the potential of generative adversarial networks in movement pattern research.

Main Methods:

• Employed a conditional cycle-consistent generative adversarial network (CycleGAN) for pairwise movement data transformation.
• Generated artificial movement data (walking, running, handwriting) from original datasets.
• Utilized a support vector machine (SVM) for classification to test the identifiability of individuals from generated data.

Main Results:

• Successfully learned pairwise transformations between vertical ground reaction force (walking/running) and vertical pen pressure (handwriting).
• Achieved high classification F1-scores, ranging from 46.8% (generated handwriting from walking) to 98.9% (generated walking from running).
• Demonstrated the identification of cross-movement individual patterns, confirming shared characteristics.

Conclusions:

• Individual movement patterns exhibit commonalities across distinct activities like walking, running, and handwriting.
• The presented methodology enables cross-movement analysis and artificial data generation.
• This approach holds potential for advancing human movement analysis and creating larger datasets for research.