A Unified Local-Global Feature Extraction Network for Human Gait Recognition Using Smartphone Sensors.

Area of Science:

• Biometrics and Human-Computer Interaction
• Machine Learning and Deep Learning

Background:

• Smartphone-based gait recognition offers unique biometric identification using inertial sensor data.
• Existing methods fail to robustly handle gait variations caused by factors like load-carrying, footwear, or clothing.

Purpose of the Study:

• To develop a novel deep learning architecture for enhanced gait recognition accuracy.
• To address the challenge of covariate factors affecting gait data by adaptively weighting features.

Main Methods:

• Proposed a weighted multi-scale Convolutional Neural Network (WMsCNN) architecture.
• Introduced a weight update sub-network (Ws) to adjust feature importance, reducing sensitivity to covariate factors.
• Integrated a fusion module to combine local and global features for improved classification.

Main Results:

• The WMsCNN model demonstrated superior performance compared to state-of-the-art deep learning approaches.
• Experiments conducted on four benchmark datasets validated the effectiveness of the proposed method.
• The weight update technique successfully mitigated the impact of covariate factors on recognition accuracy.

Conclusions:

• The proposed WMsCNN architecture effectively extracts local to global features for robust gait recognition.
• Adaptive feature weighting is a promising strategy to overcome covariate challenges in smartphone-based biometrics.
• This approach significantly boosts recognition accuracy, outperforming existing methods.