Sign in Subscribe

[Review] A Public Domain Dataset for Human Activity Recognition Using Smartphones

Last updated on 

For Jiahong’s PhD Journal
Primary source: Anguita, Ghio, Oneto, Parra, Reyes-Ortiz (ESANN 2013).

Summary

This paper introduces a foundational Human Activity Recognition (HAR) dataset collected with a single, waist-mounted smartphone (Samsung Galaxy S II) from 30 adult volunteers (19–48 years) performing six activities of daily living (ADL): walking, walking upstairs, walking downstairs, sitting, standing, laying down. The dataset—explicitly released into the public domain and mirrored at the UCI Machine Learning Repository—provides both raw accelerometer/gyroscope traces and engineered feature vectors. Using a multiclass SVM (one-vs-all, Gaussian kernel), the authors report ~96% overall accuracy on a held-out test set, with most confusion limited to the sitting vs. standing pair. This work helped standardize smartphone-based HAR benchmarking by pairing clear acquisition protocol, transparent preprocessing, and a complete feature design.

Why This Paper Matters

  1. Benchmarking: It supplies a rigorously documented, smartphone-centric dataset that many later HAR papers adopt as a baseline.
  2. Public Domain: The authors place the dataset in the public domain, lowering barriers for education, reproducibility, and derivative research.
  3. Methodological clarity: The paper details sensor placement, sampling, filtering, windowing, feature engineering, and evaluation splits, making replication straightforward.

Data Collection & Protocol

  • Participants: 30 volunteers (19–48 years).
  • Device & placement: A Samsung Galaxy S II worn at the waist. Each participant completed the protocol twice—once with fixed placement on the left belt side and once positioned by the user.
  • Activities (6 ADL): standing, sitting, laying down, walking, walking downstairs, walking upstairs.
  • Sampling rate: 50 Hz for both triaxial linear acceleration and angular velocity.
  • Acquisition environment: Lab conditions with encouragement to act naturally; 5-second rest gaps were inserted to aid segmentation and ground truthing.

Signal Processing Pipeline

  • Denoising: Median filter followed by a 3rd-order low-pass Butterworth filter (cutoff 20 Hz), justified because ~99% of human motion energy lies below 15 Hz.
  • Gravity separation: Acceleration was decomposed into body acceleration and gravity using an additional low-pass filter (0.3 Hz corner) to isolate the quasi-static gravitational component.
  • Derived signals: Time-domain magnitudes and time derivatives (e.g., jerk (da/dt), angular acceleration (d\omega/dt)). 17 total signals are considered across time and frequency domains (e.g., body/grav accel, jerks, angular speed/acc, and their magnitudes).
  • Windowing: 2.56 s sliding windows with 50% overlap (128 samples per window at 50 Hz), sized to capture ≥ one full walking cycle, including slower cadences (e.g., some elderly). FFT features are computed on each window (power-of-two length).

Feature Engineering

From each window the authors extract a comprehensive 561-dimensional feature vector. Measures include:

  • Statistical: mean, std, median absolute deviation, min/max, IQR, entropy, correlation.
  • Signal/energy: signal magnitude area (SMA), overall energy, band energy, mean frequency, dominant frequency index.
  • Shape: skewness, kurtosis.
  • Temporal models: autoregressive coefficients.
  • Orientation: angles between vector summaries (e.g., mean body acceleration vs. gravity axis).
    These are computed in both time and frequency domains across the 17 signals.

Dataset Packaging & Split

  • Two forms: (a) raw sensor streams, (b) precomputed feature vectors per window.
  • Partitioning: A random 70%/30% split into training and test sets.
  • Distribution: Hosted as “Human Activity Recognition Using Smartphones” at the UCI ML Repository with public domain terms noted in the included README (per paper).

Modeling & Results

  • Classifier: Multiclass SVM via one-vs-all with Gaussian (RBF) kernels; 10-fold cross-validation for hyperparameter selection.
  • Test set size: 2,947 windows. Overall accuracy ≈ 96%.
  • Per-class behavior:
    • Dynamic activities (walking, stairs up/down) achieve very high precision/recall (≈96%–100%).
    • Sitting vs. standing shows the main confusion; sitting recall ≈ 88% due to similar kinematics in a waist-mounted device.
  • Context vs. prior work: Results are comparable to special-purpose sensor setups (≈90–96%), supporting smartphones as viable, unobtrusive HAR platforms. Including gyroscope-driven features improved performance ~7% over a prior acceleration-only dataset from the authors.

Strengths

  1. Public domain release: Maximizes reusability in teaching, benchmarking, and open science.
  2. Complete pipeline transparency: Clear details on filtering, windowing, features, and splits.
  3. Balanced activity set: Mix of dynamic and static ADL supports both movement and posture recognition research.

Limitations & Open Challenges

  • Confusion in static postures: Waist-mounted placement makes sitting vs. standing harder to distinguish; richer orientation cues or alternative placements may help.
  • Single-device, single-placement bias: Findings may not fully generalize to pocket, handheld, or wrist usage without adaptation.
  • Controlled environment: Although “freely performed,” data were collected in lab settings; truly in-the-wild variability (device motion, carry modes, surfaces) is limited.
  • Class scope: Only six ADL; many real-world activities (cycling, running, household tasks) are out of scope.

Reproducibility & Reuse Notes

  • Signal processing: Re-implement the median and 3rd-order low-pass Butterworth (20 Hz) and gravity separation (0.3 Hz cutoff) to match the feature distributions reported.
  • Windowing: Use 2.56 s windows with 50% overlap (128 samples at 50 Hz). Maintain identical overlap to compare against reference metrics.
  • Feature parity: Recreate the 561-feature set (statistical, frequency, AR, angle, band-energy) to ensure like-for-like evaluation.
  • Baseline model: Start with OVA SVM (RBF); grid-search via 10-fold CV. Add modern variants (e.g., tree ensembles, CNN/LSTM on raw windows) for comparative studies.

Implications for Current HAR Research

  • Smartphone-as-sensor: This work validates commodity devices as accurate, low-friction HAR platforms—especially for dynamic activities.
  • Feature-rich classical ML vs. deep learning: The strong SVM performance with careful feature design sets a high classical baseline. Modern deep models should demonstrate clear, statistically robust gains to justify higher complexity and on-device costs.
  • Posture disambiguation: Persistent confusion among static postures suggests integrating gyroscope-centric features, device orientation estimation, or multi-position sensing (e.g., pocket + waist) to close the gap.

Future Directions (Inspired by the Paper)

  1. Carry-mode robustness: Benchmark across belt, pocket, hand, and bag placements; learn placement-invariant features or apply domain adaptation.
  2. On-device inference: Explore model compression (quantization, pruning) for real-time inference with minimal battery impact.
  3. Expanded label sets: Include transitions (sit↔stand), household tasks, transport modes, and health-relevant micro-activities.
  4. Multi-sensor fusion: Combine inertial with barometer, magnetometer, or environmental sensors for richer context.
  • The authors state the dataset is released to the public domain and hosted at the UCI Machine Learning Repository; the distribution includes a README describing usage terms. For this review, all descriptions are paraphrased from the cited paper, and no extended verbatim text is reproduced. Please acknowledge the original authors and the ESANN 2013 proceedings in any derivative datasets or publications.

Full Reference (Primary Source)

Anguita, D., Ghio, A., Oneto, L., Parra, X., & Reyes-Ortiz, J. L. (2013). A Public Domain Dataset for Human Activity Recognition Using Smartphones. ESANN 2013, Bruges, Belgium. (Dataset and methods as summarized above.)

Suggested Citation

Anguita, D., Ghio, A., Oneto, L., Parra, X., & Reyes-Ortiz, J. L. (2013). A Public Domain Dataset for Human Activity Recognition Using Smartphones. Proceedings of ESANN 2013. Public-domain dataset available via UCI Repository.

Note: All factual claims, figures, and protocol details in this review are drawn directly from the uploaded paper.