A Systematic Evaluation of Feature Encoding Techniques for Gait Analysis Using Multimodal Sensory Data.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Signal Processing

Background:

• Gait analysis relies on extracting features from multimodal time series sensory data.
• Wearable sensors like inertial measurement units (IMUs) are increasingly used for collecting kinematic and kinetic data.
• Effective feature encoding is critical for accurate human locomotion representation.

Purpose of the Study:

• To systematically assess various feature extraction techniques for gait analysis.
• To present and evaluate three distinct feature encoding methods for multimodal time series sensory data.
• To introduce novel deep learning models for automatic feature extraction in gait analysis.

Main Methods:

• Handcrafted feature extraction from raw sensory data.
• Bag-of-Visual-Words model with locality-constrained linear encoding (LLC).
• Two end-to-end deep learning models for automatic feature learning.

Main Results:

• Experimental evaluation on four large sensory datasets.
• Comparison with state-of-the-art methods demonstrating computational efficiency and high efficacy.
• Robustness evaluation for recognizing human daily activities.

Conclusions:

• The proposed feature encoding methods are computationally efficient and highly effective for gait analysis.
• Deep learning models offer automatic feature extraction, improving accuracy and efficiency.
• The study introduces a new dataset for gait pattern analysis using IMU sensors.