Analysis of Spatio-Temporal Representations for Robust Footstep Recognition with Deep Residual Neural Networks.

Area of Science:

• Biometrics
• Artificial Intelligence
• Pattern Recognition

Background:

• Human footsteps offer unique behavioral patterns suitable for biometric identification.
• Existing biometric systems may have limitations in certain environments or with specific data types.

Purpose of the Study:

• To develop advanced computational models for automatic biometric verification using spatio-temporal footstep representations.
• To create an artificial intelligence (AI) system capable of distinguishing subtle footstep variations between legitimate users and impostors.

Main Methods:

• Utilized floor-only sensor data to create spatio-temporal footstep representations.
• Developed and validated AI models on the largest footstep database to date (nearly 20,000 signals, >120 users).
• Evaluated performance across three security scenarios with varying training data availability: airports, workspaces, and homes.

Main Results:

• Achieved state-of-the-art footstep recognition rates.
• Reported an optimal equal error rate (EER) of 0.7%, a 371% improvement over previous methods.
• Deep residual neural networks demonstrated effective clustering of client footstep data, offering insights into feature learning.

Conclusions:

• Spatio-temporal footstep analysis using AI provides a robust method for biometric verification.
• The proposed models significantly enhance the reliability and accuracy of footstep-based biometric systems.
• The findings highlight the potential of AI in analyzing subtle human behavioral patterns for security applications.