Human Gait Recognition: A Single Stream Optimal Deep Learning Features Fusion.

Area of Science:

• Biometrics
• Computer Vision
• Artificial Intelligence

Background:

• Human Gait Recognition (HGR) is a key biometric for security, but faces challenges like varying clothing, carrying items, and different walking surfaces.
• Existing HGR techniques, especially traditional ones, struggle with large datasets and identification across diverse viewpoints.
• Deep learning methods offer potential but require effective feature selection and fusion for optimal performance.

Purpose of the Study:

• To develop an advanced deep learning framework for Human Gait Recognition (HGR).
• To address challenges in HGR, including viewpoint variations and dataset scalability.
• To improve the accuracy and efficiency of biometric identification using gait patterns.

Main Methods:

• A novel framework combining data augmentation (three flip operations), deep feature extraction (Inception-ResNet-V2, NASNet Mobile), and optimized feature selection (modified whale optimization algorithm).
• Feature fusion using the modified mean absolute deviation extended serial fusion (MDeSF) approach.
• Classification using multiple algorithms on the CASIA B gait dataset.

Main Results:

• The proposed framework achieved an average accuracy of 89.0% on the CASIA B dataset.
• Demonstrated significant improvements in accuracy, recall rate, and computational time compared to existing HGR techniques.
• Validated the effectiveness of deep learning with optimized feature selection and fusion for robust gait recognition.

Conclusions:

• The developed deep learning framework offers a more accurate and efficient solution for Human Gait Recognition.
• The integration of advanced feature extraction, selection, and fusion techniques is crucial for overcoming HGR challenges.
• This approach shows promise for real-world security applications requiring reliable individual identification.