GaitCSF: Multi-Modal Gait Recognition Network Based on Channel Shuffle Regulation and Spatial-Frequency Joint...

Area of Science:

• Biometrics and Human-Computer Interaction
• Computer Vision and Machine Learning

Background:

• Gait recognition offers non-contact, long-distance identification but struggles with real-world variations (viewpoint, clothing, occlusion, illumination) due to single-modal data limitations.
• Existing methods lack sufficient feature expression, hindering robust performance in complex, uncooperative scenarios.

Purpose of the Study:

• To develop a multi-modal gait recognition network (GaitCSF) that overcomes the limitations of single-modal approaches.
• To enhance feature representation and improve the accuracy and robustness of gait recognition systems.

Main Methods:

• Integration of two complementary modalities: silhouette data and heatmap data.
• A channel shuffle-based feature selective regulation module for cross-channel information interaction and feature enhancement.
• A spatial-frequency joint learning module utilizing Fast Fourier Transform for capturing periodic patterns and long-range dependencies.

Main Results:

• The proposed GaitCSF model achieved significant performance improvements on GREW, Gait3D, and SUSTech1k datasets.
• Demonstrated breakthrough performance compared to traditional single-modal gait recognition methods.
• Validated the effectiveness of multi-modal data integration and the proposed learning modules.

Conclusions:

• The GaitCSF model significantly enhances gait recognition performance and robustness.
• Multi-modal data fusion and spatial-frequency joint learning are effective strategies for addressing real-world challenges in gait recognition.
• The research has significant implications for practical applications requiring reliable, non-contact identification.