Optimizing Representations of Multiple Simultaneous Attributes for Gait Generation Using Deep Learning.

Area of Science:

• Computer Vision
• Biomechanical Engineering
• Machine Learning

Background:

• Gait variations are influenced by numerous factors like age, athleticism, and mood, which are difficult to quantify.
• Manual generation of diverse gait samples is challenging and often relies on simplistic, handcrafted rules.
• Existing methods struggle to capture and replicate nuanced attributes of human locomotion.

Purpose of the Study:

• To develop neural network architectures capable of learning and representing complex, hard-to-quantify gait attributes.
• To generate synthetic gait trajectories that express a custom mix of desired attributes, such as individual style and walking speed.
• To explore methods for improving the realism and diversity of synthesized human motion data.

Main Methods:

• Utilized neural network architectures to learn attribute representations from data.
• Employed cost function design and latent space regularization techniques, individually or combined.
• Integrated machine learning classifiers for both evaluating synthetic gait quality and improving training processes.

Main Results:

• Successfully demonstrated the generation of synthetic gait expressing individual style and walking speed.
• Showcased the effectiveness of cost function design and latent space regularization.
• Validated synthetic gait quality by using classifiers; a gait fooling a classifier was considered a good example.

Conclusions:

• Neural networks can effectively learn and generate complex gait attributes.
• Machine learning classifiers serve as valuable tools for both evaluating and enhancing synthetic gait generation.
• The proposed method offers a robust approach for creating diverse and realistic human motion data.