A Comprehensive Study on Cross-View Gait Based Human Identification with Deep CNNs.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Human identification based on gait patterns is a challenging biometric task.
• Existing methods often struggle with variations in viewing angles and walking conditions.

Purpose of the Study:

• To develop a novel gait recognition method using deep convolutional neural networks (CNNs) and similarity learning.
• To evaluate the proposed method's effectiveness across various challenging scenarios, including cross-view and cross-walking conditions.

Main Methods:

• Utilized deep CNNs for learning discriminative gait features from labeled multi-view walking videos.
• Employed similarity learning to enhance the recognition of subtle gait pattern changes indicative of identity.
• Conducted extensive empirical evaluations on CASIA-B, OU-ISIR, and USF gait datasets.

Main Results:

• Achieved state-of-the-art performance on the CASIA-B dataset, with recognition rates reaching 94% for large cross-view angles (≥36 degrees), significantly outperforming previous methods.
• Demonstrated excellent generalization on the large-scale OU-ISIR dataset, achieving >98% accuracy in identical view and >91% in cross-view scenarios.
• Outperformed existing methods on the USF dataset, which features real-world outdoor gait sequences.

Conclusions:

• Deep CNN-based similarity learning offers a highly effective approach for gait-based human identification.
• The proposed method shows robust performance across diverse datasets and challenging conditions, highlighting its potential for practical applications.