Feature Learning Networks for Floor Sensor-based Gait Recognition.

Area of Science:

• Computer Science
• Biometrics
• Pattern Recognition

Background:

• Deep learning (DL) models typically require large datasets for high accuracy in image classification.
• Limited data availability poses a significant challenge for DL model performance in many real-world applications.
• Traditional feature extraction methods may not fully leverage the capabilities of deep learning frameworks.

Purpose of the Study:

• To evaluate the effectiveness of feature learning networks compared to traditional methods for pressure-based footstep recognition with limited data.
• To investigate the performance of various traditional feature extraction techniques and their corresponding DL network counterparts.
• To identify the most effective feature learning network for person authentication in low-data scenarios.

Main Methods:

• Compared traditional methods: Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT), Independent Component Analysis (ICA), and Principal Component Analysis (PCA).
• Evaluated DL feature learning networks: ScatNet, DCTNet, ICANet, and PCANet, built on a Convolutional Neural Network (CNN) framework.
• Assessed performance in pressure-based footstep recognition for person authentication using limited sample sizes.

Main Results:

• Feature learning networks achieved significantly higher average accuracy (90.6%) than conventional methods (79.7%) (p < 0.05).
• PCANet demonstrated the best verification performance among all tested feature networks, reaching 92.2% accuracy.
• The study confirmed the superior performance of integrated feature learning networks over standalone traditional techniques.

Conclusions:

• Feature learning networks offer a promising and effective solution for gait recognition and person authentication in scenarios with limited training data.
• These methods are particularly relevant for security access applications like workspace environments and border control.
• The integration of traditional feature extraction with DL frameworks enhances model robustness and accuracy when data is scarce.