SPD-Net: A semantic partitioned transformer with dynamic graph network for improved skeleton-based gait recognition.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Gait recognition is a key biometric modality, but existing silhouette-based methods struggle with variations.
• Model-based methods use skeleton data but often miss semantic joint relationships.
• Transformer models capture long-range dependencies but are computationally expensive.

Purpose of the Study:

• To develop a robust and computationally efficient gait recognition system.
• To enhance the representation of gait features by modeling complex joint relationships.
• To overcome limitations of existing silhouette-based and model-based gait recognition techniques.

Main Methods:

• Proposed Semantic Partitioned transformer with Dynamic Graph Network (SPD-Net).
• Integrated Dynamic Graph Convolutional Network (DGCN) for spatial correlations, Temporal Convolutional Network (TCN) for temporal dependencies, and Semantic Partitioned Multi-head Self-Attention (SP-MSA) for focused feature extraction.
• Introduced a Joint-Part Mapping (JPM) module for hierarchical joint relationship analysis.

Main Results:

• SPD-Net demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• Achieved improved robustness and accuracy in diverse gait recognition scenarios.
• Significantly reduced computational complexity while preserving critical gait patterns.

Conclusions:

• SPD-Net offers a robust and efficient solution for gait recognition.
• The proposed semantic partitioning and dynamic graph network effectively capture complex gait dynamics.
• This approach advances the field of biometric identification through enhanced human motion analysis.