Recognition of Fine-Grained Walking Patterns Using a Smartwatch with Deep Attentive Neural Networks

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Human walking patterns vary significantly based on context and activity (e.g., smartphone use, icy conditions).
• Understanding these diverse walking styles can provide valuable contextual information for users.
• Previous methods lacked the ability to dynamically recognize a wide range of everyday walking variations.

Purpose of the Study:

• To develop a machine learning model for recognizing 18 distinct everyday walking styles.
• To create a practical, smartwatch-based wearable system for capturing and analyzing hand motion data.
• To investigate the effectiveness of deep learning algorithms in classifying walking patterns from multivariate time-series (MTS) signals.

Main Methods:

• A wearable system was developed using a commercial smartwatch to capture hand motion as multivariate time-series (MTS) signals.
• A dataset of 18 different everyday walking styles was defined and collected.
• Supervised machine learning algorithms, including feature-based and deep learning approaches (e.g., attention-based recurrent neural networks), were employed to classify the MTS data.

Main Results:

• The proposed smartwatch-based system successfully recognized a variety of predefined walking patterns.
• Deep learning algorithms demonstrated high accuracy in classifying the different walking styles based on hand motion data.
• Analysis using attention-based recurrent neural networks provided insights into the contribution of different MTS signals to classification.

Conclusions:

• Smartwatch-based sensing combined with deep learning is effective for recognizing diverse everyday walking styles.
• The system offers a promising approach for context-aware applications and personalized user feedback.
• Further research can explore more complex walking scenarios and refine the understanding of motion signal contributions.