A wearable IMU-based framework for daily physical activity recognition and energy expenditure level classification in...

Insufficient physical activity and prolonged sedentary behavior remain common among university students, yet conventional self-reports suffer from recall bias. The key idea of this research is to establish a unified, single-sensor framework that bridges engineering-based activity recognition with public health-oriented energy assessment by mapping raw inertial signals directly to metabolic intensity levels.

This study developed and evaluated a waist-worn IMU-based framework for daily physical activity recognition and energy expenditure level classification. A purpose-built dataset was collected from 96 healthy university students (mean age 20.4 ± 1.7 years), yielding 106,376 valid signal windows across six representative activities: sitting, standing, level walking, stair ascent, stair descent, and jogging. Triaxial acceleration and angular velocity signals were processed through filtering, normalization, and sliding-window segmentation.

Under leave-one-subject-out (LOSO) validation, the best-performing CNN-LSTM model achieved an accuracy of 93.8% (macro-F1: 0.932) for activity recognition and 89.6% (macro-F1: 0.883) for three-level energy expenditure classification.

The primary contribution lies in validating single-sensor feasibility within a specific campus movement ecology, demonstrating that the system can simultaneously deliver behavioral and metabolic insights without post hoc label mapping. This study fills a critical knowledge gap by quantifying the accuracy trade-offs of a joint pipeline under rigorous subject-independent conditions, providing a feasible basis for large-scale campus health surveillance.