Assessing Impact of Sensors and Feature Selection in Smart-Insole-Based Human Activity Recognition

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Human Activity Recognition

Background:

• Human Activity Recognition (HAR) systems are vital for healthcare and fitness.
• Wearable technologies aim to minimize user impact during daily activities.
• Smart insoles offer a discreet approach to HAR.

Purpose of the Study:

• To develop and evaluate an improved smart insole-based HAR system.
• To investigate the influence of data segmentation, sensor types, and feature selection on HAR performance.
• To optimize the HAR system for recognizing six distinct ambulation activities.

Main Methods:

• Utilized a Support Vector Machine (SVM) for activity classification.
• Optimized data segmentation using a 10s sliding window with 50% overlap.
• Assessed the contribution of inertial and pressure sensors, applying feature selection to reduce feature count from 272 to 227.

Main Results:

• Optimized data segmentation significantly impacted classification accuracy.
• Inertial sensors proved effective for dynamic activities; pressure sensors excelled in stationary activities.
• Combining both sensor types yielded the highest accuracy at 94.66%.

Conclusions:

• The proposed smart insole system demonstrates high accuracy in HAR.
• Sensor fusion (inertial and pressure) is crucial for comprehensive activity recognition.
• The study highlights the importance of optimized data segmentation and feature selection for robust HAR.