Wearable Inertial Sensor Dataset for Human Activity Recognition Coverage Analysis

Overview

This dataset contains anonymized inertial sensor data collected from wearable devices during human activity recognition (HAR) validation studies. The data consists of raw 3-axis accelerometer and 3-axis gyroscope readings captured during various daily activities.

Dataset Characteristics
– Sensor Type: Inertial Measurement Unit (IMU) – 3-axis accelerometer and 3-axis gyroscope
– Sampling Rate: Fixed-rate IMU sampling as defined by the commercial wearable device firmware
– Activities: 12 activity classes including walking, stairs ascent, stairs descent, forward fall, backward fall, sitting, standing, and additional daily activities
– Participants: Data were collected from multiple participants; the exact number is not disclosed as the dataset is fully anonymized and provided by a commercial partner
– Total Windows: 1,674 time-series windows
– Window Size: 5 seconds (fixed-length windows)
– Device: Generic wrist-worn wearable device
– Data Format: Excel (.xls) files, one per recording session

Data Structure
Each file contains raw sensor readings with the following columns:
– accel-X: X-axis acceleration (raw sensor units)
– accel-Y: Y-axis acceleration (raw sensor units)
– accel-Z: Z-axis acceleration (raw sensor units)
– Gyro-X: X-axis angular velocity (raw sensor units)
– Gyro-Y: Y-axis angular velocity (raw sensor units)
– Gyro-Z: Z-axis angular velocity (raw sensor units)

File Naming: Files are labeled with random pseudonymous identifiers (e.g., “participant1.xls”) for organizational purposes. These labels do not correspond to actual participant identities and no linking key exists.

Anonymization
This dataset is fully anonymized and contains NO personally identifiable information:
– No names, contact information, or demographics
– No device serial numbers or identifying metadata
– No temporal or location markers
– Random pseudonymous file labels with no linking key
– Only raw sensor measurements included

The data collection entity retained no linking information between file labels and participant identities.

Use Cases
This dataset is suitable for:
– Human activity recognition algorithm development
– Coverage analysis and data blindness studies
– Wearable sensor signal processing research
– Edge AI and TinyML model validation
– Generalization and robustness testing

Related Publication
This dataset supports the research presented in:

Pal, B., Bhattacharya, S., & Singh, M. (2025). arXiv:2604.05057.

The publication introduces a mathematical framework for measuring coverage blindness in wearable HAR systems and uses this dataset to demo