Gait Recognition by Combining the Long-Short-Term Attention Network and Personal Physiological Features

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Gait recognition performance is limited in real-world scenarios with sparse data.
• Existing methods struggle with variations in walking conditions and limited image availability.

Purpose of the Study:

• To develop an advanced gait recognition framework that integrates temporal attention mechanisms and physiological features.
• To improve gait recognition accuracy using limited visual information from monocular images.

Main Methods:

• A novel framework fusing global long-term attention (GLTA) and local short-term attention (LSTA) on silhouette sequences.
• A method for calculating personal static and dynamic physiological features from a single monocular image.
• A physiological feature extraction (PFE) network concatenating physiological data with silhouette information.

Main Results:

• Achieved a mean accuracy of rank-1 up to 89.6% on the CASIA-B dataset under various walking conditions.
• Demonstrated a 2.4% higher mean accuracy of rank-1 on the Multi-state Gait dataset compared to existing methods, particularly with different clothing.
• Validated the effectiveness and efficiency of the proposed method on benchmark datasets.

Conclusions:

• The proposed framework effectively enhances gait recognition accuracy by combining temporal attention and physiological features.
• The method shows significant promise for real-world gait recognition applications with limited data.
• Physiological feature extraction from monocular images offers a valuable addition to silhouette-based gait analysis.