IMU-Based Gait Recognition Using Convolutional Neural Networks and Multi-Sensor Fusion.

Area of Science:

• Biometrics
• Machine Learning
• Signal Processing

Background:

• Wearable sensors generate valuable human motion data for individual identification via gait patterns.
• Manual feature extraction for gait analysis is subjective and error-prone, limiting model generalization.
• Deep learning offers potential for automated feature extraction from complex motion data.

Purpose of the Study:

• To propose a novel human gait identification approach using time-frequency (TF) expansion and deep convolutional neural networks (DCNNs).
• To investigate the impact of sensor location on gait identification performance.
• To enhance gait identification accuracy through early and late multi-sensor fusion techniques.

Main Methods:

• Human gait cycles were expanded into time-frequency (TF) representations to capture joint spectral and temporal patterns.
• A deep convolutional neural network (DCNN) was designed to learn discriminative spectro-temporal features for gait identification.
• Raw motion data from five inertial measurement units (IMUs) were collected, and early/late multi-sensor fusion methods, including Minimum Error Score Fusion (MESF), were implemented.

Main Results:

• A single IMU with the 2D TF-DCNN achieved 91% subject identification accuracy.
• Early sensor fusion improved accuracy to 93.36%.
• Late sensor fusion, specifically MESF, achieved the highest accuracy of 97.06%.

Conclusions:

• The proposed 2D TF-DCNN approach effectively extracts discriminative features for human gait identification.
• Multi-sensor fusion significantly enhances the generalization performance and accuracy of gait identification systems.
• The study demonstrates the potential of deep learning and sensor fusion for robust biometric identification using wearable motion data.