Personal Identification Using Gait Spectrograms and Deep Convolutional Neural Networks.

Area of Science:

• Biometrics
• Human-Computer Interaction
• Wearable Technology

Background:

• Gait is a unique behavioral trait suitable for biometrics, offering unobtrusive, remote monitoring.
• Wearable inertial sensors enable accurate and cost-effective measurement of human motion.
• Existing biometric methods often require explicit subject involvement and physical contact.

Purpose of the Study:

• To propose and validate an approach for personal identification using kinematic gait data.
• To leverage wearable inertial sensors for reliable gait analysis.
• To assess the effectiveness of spectrographic gait representations for identification.

Main Methods:

• Collected 3-axis acceleration and angular velocity data using inertial measurement units on the feet, shanks, thighs, and pelvis.
• Applied time-frequency analysis to generate gait spectrograms from lower body movement signals.
• Utilized deep convolutional neural networks for classification with 4-fold cross-validation.

Main Results:

• Achieved 99.69% identification accuracy using spectrograms from foot, shank, thigh, and pelvic sensors.
• Attained 96.89% accuracy using only foot spectrograms.
• Demonstrated high reliability in personal identification based on gait patterns.

Conclusions:

• Kinematic and spectrographic gait analysis using wearable sensors is feasible for reliable personal identification.
• This approach shows significant potential for advancing behavior-based biometric technologies.
• Gait biometrics can be implemented unobtrusively, enhancing user convenience and privacy.