A Multi-Label Based Physical Activity Recognition via Cascade Classifier.

Area of Science:

• * Sensor-based human activity recognition
• * Machine learning for health and fitness

Background:

• * Physical activity recognition (PAR) is crucial for medical rehabilitation and fitness management.
• * Current methods struggle with recognizing complex, free-living physical activities.
• * Existing PAR models often rely on single-label classification systems.

Purpose of the Study:

• * To develop an advanced PAR system capable of recognizing complex physical activities in free-living conditions.
• * To introduce a novel multi-label classification approach for enhanced PAR accuracy.
• * To improve the generalization performance of sensor-based activity recognition models.

Main Methods:

• * Proposed a Cascade Classifier on Multi-label (CCM) structure for sensor-based PAR.
• * Utilized a two-tiered labeling system: activity intensity first, then activity type.
• * Implemented and compared the CCM approach with Random Forest (RF), Sequential Minimal Optimization (SMO), and K Nearest Neighbors (KNN) algorithms.
• * Collected data from 110 participants for experimental validation.

Main Results:

• * The proposed RF-CCM classifier achieved 93.94% accuracy, significantly outperforming the non-CCM system (87.93%).
• * The CCM system demonstrated superior recognition accuracy for ten distinct physical activities.
• * The novel CCM system showed improved generalization performance compared to conventional methods.

Conclusions:

• * The Cascade Classifier on Multi-label (CCM) system offers a more effective and stable solution for sensor-based physical activity recognition.
• * This multi-label approach enhances the ability to recognize complex, real-world activities.
• * The findings suggest CCM as a promising advancement for PAR in health and fitness applications.