A comparison of feature extraction methods for the classification of dynamic activities from accelerometer data

Area of Science:

• Biomedical Engineering
• Human Activity Recognition
• Signal Processing

Background:

• Increasing sedentary lifestyles necessitate accurate human activity monitoring for healthcare.
• Existing methods for activity classification from sensor data vary in feature extraction and classification schemes.

Purpose of the Study:

• To compare the effectiveness of 14 feature extraction methods for human activity classification using accelerometer signals.
• To evaluate performance across different activity types, accelerometer placements, and subject-based cross-validation.

Main Methods:

• Extracted features using wavelet transform, time-domain, and frequency-domain characteristics from accelerometer signals.
• Utilized two datasets: one with three activities (walking, stair ascent/descent), another with eight activities.
• Employed subject-based cross-validation with a nearest-neighbor classifier to assess classification accuracy.

Main Results:

• Frequency-based features demonstrated superior performance compared to wavelet transform for classifying dynamic activities in healthy subjects.
• The best feature sets achieved over 95% intersubject classification accuracy.
• Accelerometer placement influenced classification accuracy, with specific placements yielding better results.

Conclusions:

• Frequency-domain features are more effective than wavelet transforms for classifying dynamic human activities from accelerometer data.
• Robust feature selection and appropriate accelerometer placement are crucial for accurate human activity recognition.
• The findings provide valuable insights for developing advanced activity monitoring systems in healthcare.