Divide and Conquer-Based 1D CNN Human Activity Recognition Using Test Data Sharpening.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) utilizes sensor data from mobile devices to identify human actions.
• Deep learning has significantly advanced HAR performance on benchmark datasets.
• Existing research often focuses on signal denoising, with less attention to test data sharpening.

Purpose of the Study:

• To propose a novel 1D CNN-based approach for HAR.
• To enhance HAR accuracy through a two-stage classifier learning and test data sharpening.
• To investigate the impact of test data sharpening on HAR.

Main Methods:

• Developed a two-stage learning process using multiple 1D CNN models.
• Implemented a binary classifier for abstract activity recognition followed by multi-class classifiers for specific activities.
• Introduced test data sharpening during the prediction phase.

Main Results:

• The proposed approach demonstrated superior performance on two popular HAR benchmark datasets.
• Outperformed a two-stage 1D CNN-only method.
• Achieved better results compared to other state-of-the-art HAR methods.

Conclusions:

• The proposed 1D CNN approach with two-stage learning and test data sharpening is effective for HAR.
• Test data sharpening offers a promising technique for improving HAR accuracy.
• The method provides a significant advancement over existing HAR techniques.