Explainable gait recognition with prototyping encoder-decoder.

Area of Science:

• Biometrics
• Computer Science
• Human-Computer Interaction

Background:

• Human gait is a unique identifier, making gait recognition a key area in biometrics.
• Wearable devices are increasingly used for collecting gait data, but challenges remain in accuracy and interpretability.
• Prior research often relies on inertial measurement units, with limited exploration of pressure sensor data from insoles.

Purpose of the Study:

• To develop a novel gait recognition system using pressure sensor data from insoles.
• To propose a new neural network architecture less sensitive to hyper-parameter variations.
• To enhance the interpretability of neural network-based gait recognition models.

Main Methods:

• Collected gait data from 40 individuals using pressure-sensitive insoles.
• Developed a novel prototyping encoder-decoder network architecture for gait recognition.
• Integrated explainable AI tools, including sensitivity analysis (SA) and layer-wise relevance propagation (LRP), for model interpretability.

Main Results:

• The proposed network architecture demonstrated reduced sensitivity to hyper-parameter changes.
• The new module enabled analysis of input feature relevance for gait recognition.
• Accurate identification of gait phases was achieved using pressure data time series.

Conclusions:

• The developed system offers a robust and interpretable approach to individual recognition via human gait.
• The novel network architecture and explainable module advance the field of wearable-based biometrics.
• This research provides a foundation for more transparent and reliable gait recognition systems.