Leveraging Wearable Sensors for Human Daily Activity Recognition with Stacked Denoising Autoencoders

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Activity recognition is crucial in industrial and healthcare applications.
• Existing methods often struggle with recognizing transitional activities (e.g., sit-to-stand).
• Transitional activities are important for real-world applications.

Purpose of the Study:

• To propose a novel framework for recognizing static, dynamic, and transitional activities.
• To automatically extract features using deep learning, avoiding manual feature engineering.
• To address the challenge of unbalanced samples in transitional activity recognition.

Main Methods:

• Utilized stacked denoising autoencoders (SDAE) for automatic feature extraction.
• Employed a resampling technique (random oversampling) to handle imbalanced datasets.
• Collected data from 10 adults using wearable sensors (accelerometer and gyroscope) in a smart lab.
• Explored optimal sensor combinations for activity recognition.

Main Results:

• Achieved significant performance in recognizing transitional activities.
• Attained an overall accuracy of 94.88% across three activity types.
• Demonstrated the framework's feasibility and superiority through comparisons with other methods and public datasets.

Conclusions:

• The proposed SDAE-based framework effectively recognizes static, dynamic, and transitional activities.
• The method shows high accuracy, particularly for challenging transitional movements.
• The study validates the framework's effectiveness and potential for real-world activity recognition applications.