LightGBM-Based Human Action Recognition Using Sensors.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) using smartphones is gaining traction due to device portability.
• Distinguishing between similar activities (e.g., leaning vs. walking, stair ascent/descent) remains a challenge.
• Existing HAR methods struggle with accuracy for nuanced movements.

Purpose of the Study:

• To improve the accuracy and stability of HAR on smartphones.
• To address the discrimination issues between similar human activities.
• To enhance the efficiency of HAR models.

Main Methods:

• Utilized smartphone sensors (accelerometers and gyroscopes) for data acquisition.
• Developed a comprehensive feature extraction method incorporating time and frequency domains, generating over 300 features.
• Employed the LightGBM algorithm for analyzing extracted features and optimizing model parameters via grid search.
• Applied feature selection and dimensionality reduction techniques to improve model efficiency.

Main Results:

• The proposed feature extraction method significantly improved HAR accuracy.
• LightGBM outperformed Random Forest and XGBoost, achieving an accuracy rate of 94.98%.
• Grid search optimization further boosted LightGBM's prediction accuracy to 95.35%.
• Feature selection and dimensionality reduction increased model time efficiency by 70.14% without compromising accuracy.

Conclusions:

• The developed feature extraction and LightGBM-based approach effectively enhances HAR accuracy, particularly for similar activities.
• The study demonstrates a robust and efficient method for smartphone-based HAR.
• Optimized models show high prediction accuracy and improved computational efficiency.