GaitSG: Gait Recognition with SMPLs in Graph Structure.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait recognition identifies individuals by their unique walking patterns.
• Skinned multi-person linear (SMPL) models offer rich pose and shape data, outperforming silhouettes and skeletons.
• Previous methods underutilized SMPL parameters for gait recognition.

Purpose of the Study:

• To propose a novel gait recognition method, GaitSG, leveraging graph-structured SMPL parameters.
• To explore the potential of SMPL representations for enhanced gait analysis.
• To improve accuracy and efficiency in identifying individuals by their gait.

Main Methods:

• Representing SMPL models as graph nodes and applying graph convolution techniques.
• Developing a novel part graph pooling block (PGPB) to encode viewpoint information and address graph limitations.
• Utilizing human body prior knowledge in the model design.

Main Results:

• GaitSG demonstrated superior performance and faster convergence on public datasets (Gait3D, CASIA-B).
• Achieved approximately double the Rank-1 accuracy compared to the baseline SMPLGait (3D only) on Gait3D.
• Required three times fewer training iterations than the baseline on Gait3D.

Conclusions:

• GaitSG effectively models human topology and generates discriminative gait features using graph convolutions.
• The PGPB successfully encodes viewpoint information and mitigates limitations of graph structures.
• GaitSG represents a significant advancement in model-based gait recognition, offering improved accuracy and efficiency.