Plantar Pressure-Based Gait Recognition with and Without Carried Object by Convolutional Neural Network-Autoencoder...

Area of Science:

• Biometrics and Pattern Recognition
• Machine Learning in Security
• Human-Computer Interaction

Background:

• Convolutional Neural Networks (CNNs) excel in closed-set gait identification but struggle with open-set recognition of unknown individuals.
• Existing gait recognition methods require improvement for robustness against variations like carried objects.

Purpose of the Study:

• To develop a robust gait identification system using plantar pressure data capable of open-set recognition.
• To enhance security by distinguishing authorized users from unauthorized personnel.

Main Methods:

• Proposed a Convolutional Neural Network Autoencoder (CNN-AE) architecture for user classification.
• Extracted gait features from plantar pressure data using pressure-sensitive mats, focusing on foot pressure distribution and size.
• Implemented preprocessing techniques including ROI selection, feature image extraction, and data augmentation.
• Evaluated model performance under conditions with and without carried objects.

Main Results:

• Achieved high accuracy (91.2%) without carried objects and notable accuracy (85.6%) when carrying a 500g object.
• The CNN-AE model compressed foot pressure maps into feature vectors for identity determination.
• Demonstrated successful classification of authorized and unauthorized individuals.

Conclusions:

• The proposed CNN-AE architecture offers improved robustness for open-set gait recognition compared to traditional CNNs.
• Plantar pressure gait analysis with CNN-AE shows potential for enhanced security applications, even with moderate object carrying.
• Further research can explore more complex carrying scenarios and larger datasets.