Generative deep learning applied to biomechanics: A new augmentation technique for motion capture datasets

Area of Science:

• Biomechanics
• Machine Learning
• Data Science

Background:

• Deep learning models for biomechanics require large datasets, which are difficult to obtain from gait labs.
• Existing data augmentation techniques for motion capture data are limited.

Purpose of the Study:

• To present a generative adversarial network-based data augmentation approach for creating synthetic motion capture (mocap) datasets.
• To evaluate the effectiveness of synthetic data in improving the accuracy of deep learning biomechanical models.

Main Methods:

• An adversarial autoencoder architecture was developed, comprising an encoder, decoder, and discriminator.
• Synthetic mocap data (marker trajectories, ground reaction forces) were generated and compared to real data using statistical parametric mapping.
• The impact of synthetic data on direct kinematics (DK), inverse kinematics (IK), and inverse dynamics (ID) was assessed.

Main Results:

• Negligible differences were found for DK joint angles and ground reaction forces (GRFs) between real and synthetic data.
• Inverse methods (IK, ID) showed higher differences (29.2%, 35.5%) initially.
• Including IK joint angles in training improved joint moment estimation (ID: 25.7%) and reduced kinematic/GRF differences.
• The augmentation approach improved kinematics (up to 23%) and vertical GRF (11%) prediction accuracy in standard neural networks.

Conclusions:

• The proposed adversarial autoencoder effectively generates realistic synthetic mocap data.
• Synthetic data augmentation can significantly improve the performance of deep learning biomechanical models, especially when incorporating inverse kinematics.
• This technique offers a viable solution for overcoming data limitations in gait analysis and related fields.