Skip to content

UniMiB-SHAR

  • Modality: Smartphone accelerometer
  • Primary Tasks: Activity recognition, fall detection
  • Scale: 30 subjects, 17 activity classes (9 ADL + 8 falls), 11,771 samples
  • License: Research use only (University of Milano-Bicocca terms)
  • Access: http://www.sal.disco.unimib.it/technologies/unimib-shar/

Summary

UniMiB-SHAR (University of Milano-Bicocca Smartphone-based Human Activity Recognition) is a smartphone accelerometer dataset designed to benchmark both activities of daily living (ADL) and fall detection. The dataset includes 9 ADL types (e.g., standing, walking, going up/down stairs) and 8 fall types (e.g., forward fall, backward fall, syncope), collected from 30 subjects with a Samsung Galaxy Nexus I9250 placed in the trouser pocket. The inclusion of diverse fall types alongside normal activities makes it particularly relevant for elderly care and safety monitoring applications.

Reference Paper

  • Daniela Micucci, Marco Mobilio, Paolo Napoletano. "UniMiB SHAR: A Dataset for Human Activity Recognition Using Acceleration Data from Smartphones." Applied Sciences, 2017. PDF

Benchmarks & Baselines

  • SVM - Accuracy: ~73% (17 classes) — Micucci et al., 2017.
  • k-NN - Accuracy: ~71% (17 classes) — Micucci et al., 2017.
  • 1D-CNN - Accuracy: ~78% (17 classes) — reported in follow-up deep learning studies.
  • Fall detection (binary) - F1: >95% — commonly achieved by most methods.
  • Standard evaluation uses leave-one-subject-out cross-validation.

Tooling & Ecosystem

  • Official website provides MATLAB and CSV data files.
  • Data is provided as pre-segmented acceleration windows (151 samples per window).
  • Compatible with standard Python ML/DL pipelines (NumPy, scikit-learn, PyTorch).

Known Challenges

  • Accelerometer-only data (no gyroscope) limits the discriminability of certain activities.
  • Fall data is simulated (controlled falls), not real-world falls, which may differ in dynamics.
  • Significant class imbalance between ADL and fall classes.
  • Single sensor placement (trouser pocket) with fixed orientation.
  • Distinguishing between 8 fall subtypes is substantially harder than binary fall detection.

Cite

@article{micucci2017unimib,
  title   = {UniMiB SHAR: A Dataset for Human Activity Recognition Using Acceleration Data from Smartphones},
  author  = {Micucci, Daniela and Mobilio, Marco and Napoletano, Paolo},
  journal = {Applied Sciences},
  volume  = {7},
  number  = {10},
  pages   = {1101},
  year    = {2017}
}