Differentiating Between Walking and Stair Climbing Using Raw Accelerometry Data.

Area of Science:

• Biomechanics
• Human Movement Analysis
• Wearable Technology

Background:

• Wearable accelerometers capture high-frequency, unlabeled 3D time-series data for physical activity monitoring.
• Objective measurement of human physical activity is crucial for health and performance assessment.
• Differentiating between various walking modalities (e.g., level walking, stair ascent/descent) presents a challenge for current activity recognition systems.

Purpose of the Study:

• To develop and evaluate a classification method for detecting different types of walking using accelerometry data.
• To extract meaningful features from raw accelerometry data for improved activity classification.
• To assess the impact of sensor location, model parameters, and feature normalization on classification accuracy.

Main Methods:

• Feature extraction from raw accelerometry data using Fourier and wavelet transforms.
• Development of subject-specific and group-level classification models using a tree-based methodology.
• Evaluation of sensor location (e.g., ankle) and inter-subject feature normalization techniques.

Main Results:

• Subject-specific models achieved an average overall classification accuracy of 87.6% for differentiating walking activities.
• Ankle-worn accelerometers demonstrated the highest performance with an average accuracy of 90.5% at the subject-specific level.
• Group-level classification accuracy improved from 72.3% (unnormalized features) to 80.2% (normalized features).

Conclusions:

• A robust framework for feature extraction and classification of walking modalities from accelerometry data has been established.
• Ankle-mounted accelerometers offer superior performance for distinguishing walking sub-classes.
• Inter-subject feature normalization is a critical step for enhancing group-level activity recognition accuracy in wearable sensor studies.