Wearable Sensor-Based Human Activity Recognition via Two-Layer Diversity-Enhanced Multiclassifier Recognition Method.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Sensor-based human activity recognition (HAR) is crucial for applications in healthcare, fitness, and smart homes.
• Existing methods often face challenges in accurately discriminating between similar activities.
• Single wearable accelerometers offer a practical solution for HAR but require robust recognition algorithms.

Purpose of the Study:

• To propose a novel two-layer diversity-enhanced multiclassifier recognition method for single wearable accelerometer-based HAR.
• To improve the discriminative power of activity samples and the performance of multiclassifier systems.
• To enhance the accuracy and recall of human activity recognition.

Main Methods:

• Utilizing Kernel Fisher Discriminant Analysis (KFDA) for spatial transformation of training samples.
• Employing bootstrap resampling to increase dataset diversity for base classifier training.
• Developing a combined diversity measure for selecting optimal base classifiers.
• Implementing majority voting for combining selected base classifiers.

Main Results:

• Data-based diversity enhancement improved the discriminance between different activity samples.
• The proposed method generated base classifiers with diverse structures and performances.
• Achieved 92.3% accuracy and 90.7% recall in human activity recognition experiments.
• Outperformed random selection, Bagging, and Adaboost ensemble methods.

Conclusions:

• The proposed two-layer diversity-enhanced multiclassifier method significantly improves HAR performance.
• KFDA and bootstrap resampling effectively enhance data discrimination and classifier diversity.
• The approach offers a superior alternative to traditional ensemble methods for accelerometer-based HAR.