Condition-Adaptive Graph Convolution Learning for Skeleton-Based Gait Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Skeleton-based gait recognition is crucial for identifying individuals based on their walking patterns.
• Existing methods struggle with viewpoint variations and individual walking style differences.
• Uniform convolutions in current models fail to capture view-specific nuances.

Purpose of the Study:

• To develop a novel network that adapts to specific skeleton sequences and view angles for improved gait recognition.
• To overcome the limitations of uniform convolutions in handling diverse walking styles and viewpoints.
• To enhance the accuracy and robustness of skeleton-based gait recognition systems.

Main Methods:

• Proposed a condition-adaptive graph (CAG) convolution network.
• Introduced a joint-specific filter learning (JSFL) module for sequence-adaptive filters at the joint level.
• Developed a view-adaptive topology learning (VATL) module for adaptive graph topology generation.

Main Results:

• CAG dynamically adapts to individual walking styles and viewpoint changes.
• JSFL captures fine-grained, joint-level patterns for diverse spatial-temporal information.
• VATL generates adaptive graph topologies for view-specific joint correlations.
• Achieved state-of-the-art performance on CASIA-B and OU-MVLP datasets.
• Demonstrated complementary information when combined with appearance-based methods.

Conclusions:

• CAG significantly improves skeleton-based gait recognition by addressing viewpoint and style variations.
• The proposed adaptive approach offers a more robust and accurate solution for biometric identification.
• CAG's ability to provide complementary information enhances its applicability in multimodal recognition systems.