Improving Human Activity Recognition Performance by Data Fusion and Feature Engineering.

Area of Science:

• Biomedical Engineering
• Computer Science
• Sensor Technology

Background:

• Human activity recognition (HAR) is crucial for health monitoring.
• Wearable sensor fusion offers portable, accurate, and non-interventional HAR.
• Real-time HAR systems require compact and reliable designs, necessitating efficient feature selection.

Purpose of the Study:

• To propose a novel two-stage genetic algorithm-based feature selection algorithm with a fixed activation number (GFSFAN).
• To evaluate the effectiveness of GFSFAN in improving classification performance (CP) and reducing system burden for HAR.
• To assess the impact of sensor calibration and joint angle estimation on HAR accuracy.

Main Methods:

• Implemented GFSFAN on datasets featuring time, frequency, and time-frequency domain features from nine activities of daily living (ADL).
• Utilized six different classifiers to evaluate feature subsets selected by GFSFAN and other FS algorithms.
• Introduced sensor-to-segment coordinate calibration and lower-limb joint angle estimation algorithms.

Main Results:

• GFSFAN achieved good classification performance (CP) with a significantly smaller feature subset size.
• The proposed GFSFAN algorithm demonstrated superior performance compared to other FS methods.
• Both sensor calibration and joint angle estimation algorithms were shown to improve HAR classification performance.

Conclusions:

• GFSFAN is an effective feature selection method for developing compact and reliable real-time HAR systems.
• Integrating sensor calibration and joint angle estimation further enhances the accuracy of wearable sensor-based HAR.
• The findings support the application of advanced feature selection and sensor processing techniques in health-related HAR.