ST-DeepGait: A Spatiotemporal Deep Learning Model for Human Gait Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Human gait analysis involves complex spatiotemporal data from moving joints.
• Individual gait patterns serve as unique movement signatures for identification.
• Existing methods may not fully capture the intricate spatiotemporal dynamics of gait.

Purpose of the Study:

• To present ST-DeepGait, a spatiotemporal deep learning model for human gait recognition.
• To effectively featurize and classify co-movement patterns of human joints.
• To enable accurate individual identification based on unique gait signatures.

Main Methods:

• Developed ST-DeepGait, a deep learning architecture based on spatiotemporal human skeletal graphs.
• Utilized a multi-layer Recurrent Neural Network (RNN) to model gait cycles sequentially.
• Trained and evaluated the model on a novel RGB-D gait dataset with 100 subjects.

Main Results:

• Achieved over 90% recognition accuracy on the primary dataset.
• Demonstrated interpretable feature separability in a geometric latent space via qualitative evaluation.
• Attained 88% accuracy in zero-shot detection on unseen data, showcasing generalizability.

Conclusions:

• ST-DeepGait effectively recognizes human gaits by learning spatiotemporal co-movement patterns.
• The model exhibits strong performance and generalizability, validated by high accuracy rates.
• The ST-DeepGait model has potential applications beyond gait analysis, including sports and traffic pattern analysis.