Deep Neural Network-Based Gait Classification Using Wearable Inertial Sensor Data.

Area of Science:

• Biometrics
• Machine Learning
• Wearable Technology

Background:

• Human gait is a recognized behavioral biometric for identification.
• Classifying gait using wearable inertial sensors is an active research area.
• Neural networks offer powerful tools for complex pattern recognition in biological signals.

Purpose of the Study:

• To develop an effective neural network-based approach for gait classification.
• To evaluate the performance of different neural network models using inertial sensor data.
• To assess the feasibility of gait classification for personal identification and authentication.

Main Methods:

• Collected 3-axis accelerometer and gyroscope data from multiple body locations (pelvis, thighs, shanks, feet).
• Utilized data from 29 athletes, 19 normal foot participants, and 21 foot deformity patients.
• Developed and validated classifiers using fully connected neural networks and convolutional neural networks (CNNs) with gait spectrograms.

Main Results:

• A fully connected neural network achieved 93.02% accuracy in classifying the three groups.
• A CNN-based classifier reached 98.19% accuracy using gait spectrograms.
• Classification using only pelvic spectrograms with CNNs yielded 94.25% accuracy, reducing feature engineering needs.

Conclusions:

• The proposed neural network approaches demonstrate high accuracy and practicality for gait classification.
• CNNs applied to gait spectrograms show superior performance compared to traditional methods.
• Gait analysis using wearable sensors holds significant potential for biometric applications.