Human activity recognition (HAR) with wearable sensors is widely applied in health monitoring, fitness tracking, and smart environments, but the choice of sensor configuration remains a critical factor for balancing recognition performance with usability and comfort. Existing datasets often lack the full-body coverage required to systematically evaluate sensor placement strategies. We present a comprehensive dataset of 12 daily activities performed by 30 participants, recorded using 17 inertial measurement units (IMUs) distributed across the entire body. Each IMU provides tri-axial acceleration and angular velocity signals at 60 Hz, aligned within a standardized global coordinate system. The dataset further includes detailed anthropometric metadata, structured annotations of activity and effort level, and processing scripts to support feature extraction, segmentation, and baseline model training. Benchmark experiments with both machine learning and deep learning models demonstrate the usability of the dataset across multiple temporal windows and sensor subsets. This resource enables systematic evaluation of sensor layout strategies and supports the development of practical, generalizable HAR systems.
Feng et al. (Thu,) studied this question.