Gait Activity Classification on Unbalanced Data from Inertial Sensors Using Shallow and Deep Learning.

Area of Science:

• Ubiquitous Computing
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Activity recognition, particularly gait analysis, is crucial for identifying health risks linked to physical activity.
• Unbalanced datasets in gait recognition pose challenges, leading models to favor majority classes.
• Accurate classification of diverse gait activities (inclines, level ground, stairs) is essential.

Purpose of the Study:

• To evaluate and compare shallow and deep learning methods for gait activity classification.
• To investigate the impact of data treatments (original, sampled, augmented) on classification performance.
• To address the issue of unbalanced datasets in gait activity recognition.

Main Methods:

• Utilized accelerometer and gyroscope data from a large-scale public dataset.
• Implemented conventional (shallow) and deep learning classification techniques.
• Employed data augmentation by generating synthetic gait data to address class imbalance.

Main Results:

• Classifiers built with augmented data achieved superior performance.
• Deep learning models with augmented data yielded the highest F-measure (0.927 ± 0.033).
• Shallow learning models with augmented data also showed significant improvement (F-measure: 0.812 ± 0.078).

Conclusions:

• Data augmentation is an effective strategy for improving gait activity classification accuracy, especially with unbalanced datasets.
• Deep learning approaches, combined with data augmentation, offer the highest performance for recognizing diverse gait activities.
• The findings support the use of augmented data in gait recognition for more reliable health-related physical activity assessments.