A Comparison Study of Classifier Algorithms for Cross-Person Physical Activity Recognition.

Area of Science:

• Human-Computer Interaction
• Machine Learning
• Wearable Technology

Background:

• Physical activity is vital for health, with benefits including weight management and reduced chronic disease risk.
• Advances in wearable devices generate vast amounts of physical activity data, increasing the need for accurate activity recognition.
• Deep learning, utilizing artificial neural networks, excels at identifying complex patterns in large datasets for classification tasks.

Purpose of the Study:

• To compare various classification techniques for automatic cross-person activity recognition.
• To evaluate performance under different data availability scenarios.
• To integrate deep learning using Google's TensorFlow framework.

Main Methods:

• Utilized the PAMAP2 (Physical Activity Monitoring in the Ageing Population) dataset.
• Employed leave-one-subject-out (LOSO) cross-validation for cross-person prediction.
• Compared traditional classifiers with deep neural networks.

Main Results:

• High accuracies (e.g., 96%) achieved with traditional classifiers on large datasets.
• Deep neural networks outperformed other methods with drastically reduced data (0.001%), reaching 60% accuracy.
• Comparable results to full dataset usage were achieved with only ~22.67% of the data.

Conclusions:

• Deep learning models are effective for physical activity recognition, especially with limited data.
• Significant data reduction is possible without compromising statistical accuracy.
• Findings support the development of energy-efficient devices and improved data processing for activity records.